Quantum Monte Carlo simulations of solids
W. M. C. Foulkes
CMTH Group, Department of Physics, Imperial College of Science, Technology
and Medicine, Prince Consort Road, London SW7 2BZ, United Kingdom
L. Mitas
National Center for Supercomputing Applications, University of Illinois
at Urbana-Champaign, Urbana-Champaign, Illinois 61801 and Department of Physics,
North Carolina State University, Raleigh, North Carolina 27695-8202
R. J. Needs and G. Rajagopal
TCM Group, Cavendish Laboratory, Madingley Road, Cambridge CB3 0HE,
United Kingdom
(Published 5 January 2001)
This article describes the variational and fixed-node diffusion quantum Monte Carlo methods and how
they may be used to calculate the properties of many-electron systems. These stochastic
wave-function-based approaches provide a very direct treatment of quantum many-body effects and
serve as benchmarks against which other techniques may be compared. They complement the less
demanding density-functional approach by providing more accurate results and a deeper
understanding of the physics of electronic correlation in real materials. The algorithms are intrinsically
parallel, and currently available high-performance computers allow applications to systems containing
a thousand or more electrons. With these tools one can study complicated problems such as the
properties of surfaces and defects, while including electron correlation effects with high precision. The
authors provide a pedagogical overview of the techniques and describe a selection of applications to
ground and excited states of solids and clusters.
CONTENTS
I. Introduction
II. Interacting Electrons in Solids
A. The many-electron Schrödinger equation
B. Hartree-Fock theory
C. Post-Hartree-Fock methods
D. Density-functional theory
E. Quantum Monte Carlo methods
III. Monte Carlo Methods
A. Statistical foundations
B. The Metropolis algorithm
C. Variational Monte Carlo
D. Diffusion Monte Carlo
1. The imaginary-time Schrödinger equation
2. The fixed-node approximation
a. One-electron example
b. Many-electron version
c. The fixed-node variational principle
d. The tiling theorem
3. Importance sampling
IV. Trial Wave Functions
A. Introduction
B. Slater-Jastrow wave functions
C. The Slater determinant
D. The Jastrow factor
E. Spin
F. The cusp conditions
V. Selected Applications of Quantum Monte Carlo to
Ground States
A. Cohesive energies of solids
B. Phases of the electron gas
C. Static response of the electron gas
D. The relativistic electron gas
E. Exchange and correlation energies
Reviews of Modern Physics, Vol. 73, No. 1, January 2001
34
35
35
36
36
37
38
39
39
40
40
41
41
43
43
44
45
46
46
49
49
49
51
51
53
54
55
55
55
57
57
57
VI.
VII.
VIII.
IX.
X.
F. Compton scattering in Si and Li
G. Solid hydrogen
H. Clusters
I. Formation energies of silicon self-interstitials
J. Jellium surfaces
Excited States
A. Introduction
B. VMC and DMC calculations of excitations in
solids
C. Other QMC methods for excited states
Wave-Function Optimization
A. Introduction
B. The cost function
C. Numerical stability of variance minimization
D. Minimization procedures
Pseudopotentials
A. The need for pseudopotentials
B. Nonlocal pseudopotentials
C. Core-polarization potentials
D. Pseudopotentials in variational Monte Carlo
E. Pseudopotentials in diffusion Monte Carlo
F. Alternatives to nonlocal pseudopotentials
Periodic Boundary Conditions and Finite-Size
Errors
A. Introduction
B. Bloch’s theorem for many-body systems
C. Periodic boundary conditions and Coulomb
interactions
D. Coping with finite-size errors
1. Introduction
2. Finite-size correction and extrapolation
formulas
3. Choosing the simulation-cell wave vector
4. Interaction energy finite-size effects
Computational Issues
A. Representation of the single-particle orbitals
B. Evaluation of the trial wave function
C. Evaluation of the local energy
0034-6861/2001/73(1)/33(51)/$25.20
©2001 The American Physical Society
59
60
61
62
62
64
64
64
65
65
65
66
67
67
67
67
68
68
68
69
70
71
71
71
72
73
73
73
74
74
75
75
75
76
33
34
Foulkes et al.: Quantum Monte Carlo simulations of solids
D. Efficient treatment of the Coulomb interaction
E. Scaling with the number of particles
F. QMC on parallel computers
XI. Conclusions
Acknowledgments
References
76
76
77
77
79
79
I. INTRODUCTION
The past few decades have seen dramatic improvements in our ability to simulate complicated physical
systems using computers. This has led to the concept of
computer simulation as a ‘‘third way’’ of doing science,
closer to experiment than theory but complementary to
both. Like experiments, simulations produce data rather
than theories and should be judged on the quality of
those data. We aim to show that the quantum Monte
Carlo (QMC) methods described in this review are accurate and reliable and that the data and insights they
provide are often difficult or impossible to obtain in any
other way.
In studying the quantum-mechanical properties of solids, we are fortunate that the underlying physical laws
have been known for over 70 years. The difficulty arises
from the complexity of the equations that, as Dirac
(1929) famously wrote, are ‘‘much too complicated to be
soluble.’’ It hardly seems possible that one could solve
the 3N-dimensional Schrödinger equation that describes
the N interacting electrons in a solid, but this is exactly
what QMC methods allow us to do. With current parallel computers we can simulate periodic systems containing up to a thousand or so electrons, which is enough to
model many properties of solids with impressive accuracy. The quantum Monte Carlo method is now recognized as a valuable tool for calculating the physical properties of real materials.
The term ‘‘quantum Monte Carlo’’ covers several different techniques based on random sampling. The simplest of these, variational Monte Carlo (VMC), uses a
stochastic integration method to evaluate expectation
values for a chosen trial wave function. In a system of
1000 electrons the required integrals are 3000 dimensional, and for such problems Monte Carlo integration is
much more efficient than conventional quadrature
methods such as Simpson’s rule. The main drawback of
VMC is that the accuracy of the result depends entirely
on the accuracy of the trial wave function. The other
method we consider, diffusion Monte Carlo (DMC),
overcomes this limitation by using a projection technique to enhance the ground-state component of a starting trial wave function. During the last ten years it has
become clear that VMC and DMC can produce very
accurate ground-state expectation values for weakly correlated solids. They can also provide some information
about specific excited states, although not yet full excitation spectra.
Although VMC and DMC are the only methods currently used to study realistic continuum models of electrons in solids, other methods are useful in other systems. Auxiliary-field Monte Carlo (for a recent review
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
see Senatore and March, 1994) is often applied to model
Hamiltonians such as the Hubbard model, and pathintegral Monte Carlo (reviewed by Ceperley, 1995a,
1995b) is the main QMC method used to study bosonic
systems such as liquid helium. Both auxiliary-field and
path-integral QMC are in principle capable of simulating interacting electrons at finite temperature and may
one day be used to study real solids. In this review, however, we focus on the two techniques that have already
proved their worth in solids: the zero-temperature VMC
and DMC methods.
Because VMC and DMC simulations are very expensive, most calculations of the ground-state properties of
solids use less accurate methods based on HohenbergKohn-Sham density-functional theory or Hartree-Fock
(HF) theory (see, for example, Parr and Yang, 1989).
The efficiency of these techniques arises from the replacement of the electron-electron Coulomb interactions by an effective one-electron potential; this makes
the many-electron Schrödinger equation separable and
hence much easier to solve. Density-functional theory,
in particular, has been remarkably successful and is now
used in fields ranging from solid-state physics and quantum chemistry to geophysics and molecular biology.
Although density-functional theory is exact in principle, it relies in practice on approximations to the unknown exchange-correlation energy functional, which
accounts for the complicated correlated motion of the
electrons. Most density-functional calculations use either the local-density approximation or the generalized
gradient approximation, both of which work surprisingly
well under most circumstances. Obvious failures are uncommon, but include the underestimation of s-to-d
transfer energies in transition-metal atoms and difficulties in describing van der Waals bonding, both of which
have been dealt with successfully using QMC techniques. Problems also occur in covalent sp-bonded materials, as demonstrated by the QMC studies of carbon
clusters described in Sec. V. The unreliability of the
density-functional results for these clusters provides a
good demonstration of the limited predictive power of
existing approximate exchange-correlation energy functionals.
The strong electronic correlations in solids containing
d and f electrons produce a wide range of subtle effects,
including the anomalous normal-state transport properties of high-temperature superconductors and the very
large low-temperature specific-heat coefficients of
heavy-fermion materials. Correlation effects such as
these are not well described by density-functional or HF
theory and are normally studied using simplified Hamiltonians such as the Hubbard model. Given a set of
atomic positions, it is often possible to devise a model
Hamiltonian that gives an excellent account of the lowlying electronic excitations on an energy scale that may
be as low as a few degrees Kelvin. The simplicity is not
achieved without cost, however, since most model
Hamiltonians drastically approximate the ‘‘independentelectron’’ energy given so accurately by HF and densityfunctional calculations. The independent-electron con-
35
Foulkes et al.: Quantum Monte Carlo simulations of solids
tribution is normally more than 99% of the total energy,
and so model Hamiltonians rarely give accurate groundstate energies as functions of the atomic positions. This
review is concerned with studies of interatomic forces
and chemical bonding, for which the relevant energy
scales range from a few meV to several eV, and for
which continuum approaches based on the full manyelectron Schrödinger equation are more appropriate.
Continuum QMC simulations require much more
computer time than density-functional or HF calculations, but can already achieve chemical accuracy (usually defined as 1 kcal per mole, which is about 0.04 eV
per molecule) in small systems and may soon achieve it
in solids. Chemical accuracy is sufficient to address most
issues involving interatomic forces and chemical reactions, but is still an order of magnitude worse than the
accuracy required to study phenomena such as superconductivity. The computational cost of a QMC calculation increases as the cube of the number of electrons,
which makes applications to large systems feasible, and
there is no fundamental reason why the accuracy should
fall off for large systems. Quantum Monte Carlo algorithms are also quite simple to program and very well
suited to massively parallel computation.
Quantum Monte Carlo methods have already proved
themselves important in a very broad sense—they can
provide much more than just accurate numbers. A good
example is provided by Ceperley and Alder’s (1980)
simulations of the homogeneous electron gas, which underlie the accurate parametrized local-density approximations used in most density-functional calculations.
The rapid acceptance of these QMC-based approximations helped spark the 20 years of growth and development that made density-functional methods so popular
and powerful.
Our intention is that most of what follows should be
accessible to a wide audience, but no doubt some readers will find it helpful to consult other sources. Kalos and
Whitlock (1986) give a straightforward introduction to
Monte Carlo methods in general, and Hammond,
Lester, and Reynolds (1994) cover the use of QMC in
quantum chemistry. Other useful recent sources are the
review by Anderson (1999) and the books edited by
Lester (1997) and Nightingale and Umrigar (1999).
In Sec. II we give an overview of techniques used to
study the quantum mechanics of many-electron systems.
Section III discusses the statistical foundations of the
Monte Carlo approach and describes the theory behind
the VMC and DMC methods. The successes of VMC
and DMC rest on the surprising accuracy of relatively
simple trial wave functions, which are considered in Sec.
IV. In Sec. V we describe a selection of QMC simulations of the ground-state properties of solids and clusters, while in Sec. VI we highlight studies of excited
states. In Sec. VII we describe the methods used to optimize trial wave functions, and in Sec. VIII we discuss
pseudopotentials and why they are needed in QMC
simulations. Section IX looks at the issue of periodic
boundary conditions and the proper treatment of finiteRev. Mod. Phys., Vol. 73, No. 1, January 2001
size errors. In Sec. X we consider issues of computational efficiency. Section XI concludes this review.
II. INTERACTING ELECTRONS IN SOLIDS
A. The many-electron Schrödinger equation
One of the great challenges of condensed-matter
physics is to obtain accurate approximate solutions of
the many-electron Schrödinger equation. Because the
mass of an electron is so small compared to that of a
nucleus (m e /M⬇10⫺3 – 10⫺5 ), the dynamics of electrons
and nuclei in solids can, to a good approximation, be
decoupled. Within the Born-Oppenheimer approximation, the many-electron wave function and energy may
then be obtained by solving a time-independent Schrödinger equation in which the nuclear positions are considered fixed. The nonrelativistic Born-Oppenheimer
Hamiltonian takes the form
Ĥ⫽⫺
1
2
兺i ⵜ 2i ⫺ 兺i 兺␣ 兩 ri ⫺d␣ ␣兩
⫹
1
2
,
兺i 兺
j⫽i 兩 ri ⫺rj 兩
Z
1
(2.1)
where ri are the electron positions, d␣ are the nuclear
positions, and Z ␣ are the nuclear charges.1
The first quantum-mechanical studies of chemically
bonded systems appeared soon after the birth of quantum mechanics. Heitler and London (1927) calculated
the binding energy and internuclear separation of H2 by
approximating the wave function as an antisymmetrized
product of atomic 1s functions centered on the two nuclei. A different approach was taken by Hartree (1928),
Fock (1930), and Slater (1930), who argued that it is
reasonable to replace the intractable system of interacting electrons with one involving independent electrons
moving in a self-consistent field. Coulson and Fischer
(1949) showed that the Heitler-London and HartreeFock (HF) wave functions were exact in the limiting
cases of strong electron correlation and weak electron
correlation, respectively, and that the truth lay somewhere in between. It turns out that many chemically
bonded systems are weakly correlated, so that the HF
approach is normally a good starting point. This point of
view underlies the molecular orbital theory developed
by Hund (1928), Mulliken (1928), Hückel (1931a, 1931b,
1932), and others.
Calculating the electronic structure of solids poses a
formidable challenge because of the large number of
particles involved. The first steps were taken by Sommerfeld and others, who made the bold proposal that in
metals the forces on the electrons due to the ions approximately cancel the forces due to the other electrons
(Sommerfeld and Bethe, 1933), leading to the freeelectron model. The work of Bloch (1928) added the
We use Hartree atomic units, e⫽m e ⫽ប⫽4 ␲ ⑀ 0 ⫽1, for all
equations.
1
36
Foulkes et al.: Quantum Monte Carlo simulations of solids
important concepts of Bloch waves and energy bands. It
was soon realized, however, that HF theory fails to give
a proper description of metals, and hence that correlation effects have to be included from the outset.
The most important and widely used electronic structure method currently applied to solids is based on the
many-electron density-functional theory developed by
Hohenberg and Kohn (1964) and Kohn and Sham
(1965). This approach reduces the complicated manyelectron problem to an independent-electron problem
with an effective one-electron potential that depends on
the electron density. Although at first glance it appears
that density-functional theory must be an approximate
method, it is in principle capable of describing electron
correlation effects exactly. Density-functional theory has
been the main tool used to study the electronic structure
of solids for more than two decades (Jones and Gunnarsson, 1989; Parr and Yang, 1989; Dreizler and Gross,
1990), and over the last decade has also become popular
in quantum chemistry.
The Hartree-Fock and density-functional methods are
very relevant to our discussion of QMC. The QMC techniques we focus on here rely on the availability of reasonable approximations to the many-electron wave
function, which are often constructed using one-electron
orbitals obtained from HF or density-functional calculations. A brief introduction to these methods and to correlated wave-function methods is given in the following
subsections. Many of the concepts and equations will be
referred to later in the article.
where ␦ ␴ i , ␴ j ⫽1 if ␴ j ⫽ ␴ i and zero otherwise. If the determinant contains N ↑ orbitals with ␴ i ⫽↑ and N ↓ ⫽N
⫺N ↑ with ␴ i ⫽↓, it is an eigenfunction of Ŝ z with eigenvalue (N ↑ ⫺N ↓ )/2.
The exact ground-state wave function of an interacting system cannot be represented as a single Slater determinant, but a Slater determinant can nevertheless be
used as a variational trial function. If this trial function is
optimized by minimizing the expectation value of Ĥ
with respect to the orbitals ␺ i (r), one obtains the selfconsistent HF equations,
冉
冊
1
⑀ i ␺ i 共 r兲 ⫽ ⫺ ⵜ 2 ⫹V ion共 r兲 ␺ i 共 r兲
2
⫹
兺冕
⫺
兺j ␦ ␴ , ␴ 冕 dr⬘
j
dr⬘
i
兩 ␺ j 共 r⬘ 兲 兩 2
␺ 共 r兲
兩 r⫺r⬘ 兩 i
j
␺ j* 共 r⬘ 兲 ␺ i 共 r⬘ 兲
␺ j 共 r兲 ,
兩 r⫺r⬘ 兩
(2.5)
where the Lagrange multipliers ⑀ i arise from the orthonormality constraints on the single-particle orbitals. The
ionic potential is
V ion共 r兲 ⫽⫺
Z
兺␣ 兩 r⫺d␣␣兩 ,
(2.6)
as in Eq. (2.1), while the other two terms, known as the
Hartree and exchange terms, describe the electronelectron interactions. The effect of the exchange term is
to keep electrons of like spin apart and, as a result, each
electron has around it a Fermi or exchange hole containing unit positive charge.
B. Hartree-Fock theory
C. Post-Hartree-Fock methods
The HF approximation (Fock, 1930; Slater, 1930) is
the simplest theory that correctly incorporates the permutation symmetry of the wave function. The manyelectron wave function must be antisymmetric under
particle exchange, so that
⌿ 共 ...,xi ,...,xj ,..., 兲 ⫽⫺⌿ 共 ...,xj ,...,xi ,..., 兲 ,
(2.2)
where xi ⫽ 兵 ri , ␴ i 其 represents the space and spin coordinates of an electron. The antisymmetry ensures that no
two electrons can have the same set of quantum numbers and that the Pauli exclusion principle is satisfied.
The simplest wave function with the required antisymmetry is a Slater determinant,
冏
D 共 X兲 ⫽
␺ 1 共 x1 兲
␺ 1 共 x2 兲
...
␺ 1 共 xN 兲
␺ 2 共 x1 兲
␺ 2 共 x2 兲
...
␺ 2 共 xN 兲
]
]
]
]
␺ N 共 x1 兲
␺ N 共 x2 兲
...
冏
,
(2.3)
␺ N 共 xN 兲
where X⫽(x1 ,x2 ,...,xN ) is shorthand for the list of all
the electron coordinates. In most cases the singleparticle orbitals are assumed to be products of spatial
and spin factors,
␺ i 共 xj 兲 ⫽ ␺ i 共 rj 兲 ␦ ␴ i , ␴ j ,
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
(2.4)
The single-determinant HF theory includes the exchange effects arising from the antisymmetry of the
many-electron wave function, but neglects the electronic
correlations caused by the electron-electron Coulomb
repulsion. Correlation energies are a small fraction of
the total energy, but they can be very important. For
example, the correlation energy of a nitrogen molecule
amounts to only 0.5% of the total electronic energy, but
accounts for nearly 50% of the molecular binding energy.
Correlation can be included by using a linear combination of determinants in a post-Hartree-Fock method.
However, the central problem with such expansions is
that very large numbers of determinants are needed to
describe many-electron wave functions accurately.
There are two reasons for this poor convergence. The
first, which applies equally in small and large systems, is
that many determinants are needed to describe the cusplike gradient discontinuities that occur whenever two
electrons have the same position (see Sec. IV). The second is that the required number of determinants increases very rapidly with system size.
In a full configuration-interaction calculation one includes all the determinants that can be formed from the
37
Foulkes et al.: Quantum Monte Carlo simulations of solids
molecular orbitals calculated with a particular basis set.
Consider, for example, a system of N electrons occupying states chosen from a set of 2N molecular orbitals.
The total number of N-electron Slater determinants is
共 2N 兲 ! e
⬇ 2N ln N ⫽e 2N ln 2 ,
N!N!
e
electron density of the interacting system in terms of the
one-electron wave functions of the noninteracting system,
N
n 共 r兲 ⫽
2N ln 2N
(2.7)
for large N. The number of determinants therefore rises
exponentially with N. The full configuration-interaction
approach in conjunction with a good basis set can be
applied only to small systems, and the current practical
limit for highly accurate calculations is reached for small
molecules such as H2O.
One way of reducing the computational cost is to include only the most important determinants. This may
be done by considering low-energy excitations from a
reference determinant, which is normally the HF ground
state. For example, if single and double excitations are
included we have the configuration-interaction singles
and doubles method, for which the computational
cost scales as N 6 . Unfortunately, such truncated
configuration-interaction methods are not size consistent; that is, the energy does not scale linearly with the
number of electrons. For example, in a configurationinteraction singles and doubles calculation for N widely
separated hydrogen molecules, the correlation energy
increases only as 冑N. Such a method is clearly unsuitable for applications to solids. The size consistency problem can be overcome via coupled-cluster expansions
(Čı́žek, 1969). These implicitly include all excitations
from the reference determinant, although the coefficients in the expansion are approximated and the
method is nonvariational. Coupled-cluster methods are
capable of yielding highly accurate results and are direct
competitors of QMC methods for molecular calculations, but they are very expensive in large systems. For
example, the computational cost of a CCSD calculation
(coupled cluster with single and double excitations)
scales as N 6 . Hartree-Fock and post-Hartree-Fock
methods are described in a straightforward manner in
the book by Szabo and Ostlund (1989).
兺 兩 ␺ i 共 r兲 兩 2 ,
and to write the Hohenberg-Kohn energy functional in
the form
冕 ␺*
冕冕
N
1
E 关 n 共 r兲兴 ⫽⫺
2 i⫽1
兺
⫹
1
2
Computational methods based on density-functional
theory are now the most popular approach for calculating the electronic properties of solids and large molecules. Density-functional theory is a formally exact
theory based on a theorem of Hohenberg and Kohn
(1964), which states that the ground-state properties of a
many-electron system may be obtained by minimizing a
functional E 关 n(r) 兴 of the electron number density n(r).
The minimum value of the functional is the exact
ground-state energy, and the minimum is attained when
n(r) is the exact ground-state density.
Kohn and Sham (1965) introduced the idea of an auxiliary noninteracting system with the same electron density as the real system. This enabled them to express the
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
i
共 r兲 ⵜ 2 ␺ i 共 r兲 dr⫹
冕
n 共 r兲 V ion共 r兲 dr
n 共 r兲 n 共 r⬘ 兲
dr dr⬘ ⫹E xc 关 n 共 r兲兴 ,
兩 r⫺r⬘ 兩
(2.9)
where the terms on the right-hand side are the kinetic
energy of the noninteracting system with electron density n(r), the energy of interaction with the ionic potential, the Hartree energy, and the exchange-correlation
energy. Equation (2.9) may be taken as the definition of
the exchange-correlation energy functional E xc 关 n(r) 兴 ,
which is not known exactly and has to be approximated.
By minimizing the total energy functional of Eq. (2.9),
Kohn and Sham showed that one could calculate the
ground-state ␺ i , and hence the ground-state number
density and energy, by solving the self-consistent set of
equations,
冉
冊
1
⫺ ⵜ 2 ⫹V ion共 r兲 ⫹V H 共 r兲 ⫹V xc 共 r兲 ␺ i 共 r兲 ⫽ ⑀ i ␺ i 共 r兲 ,
2
(2.10)
where the Hartree potential is
V H 共 r兲 ⫽
冕
dr⬘
n 共 r⬘ 兲
兩 r⫺r⬘ 兩
(2.11)
and the exchange-correlation potential is
V xc 共 r兲 ⫽
␦ E xc 关 n 共 r兲兴
.
␦ n 共 r兲
(2.12)
The best-known approximation for E xc 关 n(r) 兴 is the
local-density approximation (LDA),
LDA
E xc
关 n 共 r兲兴 ⫽
D. Density-functional theory
(2.8)
i⫽1
冕⑀
hom
xc „n 共 r 兲 …n 共 r 兲 dr,
(2.13)
hom
(n) is the exchange-correlation energy per
where ⑀ xc
electron in a uniform electron gas of density n. The nonuniform electron gas at r is therefore treated as if it were
part of a uniform electron gas of constant density n
⫽n(r). This approximation is obviously accurate when
the electron density is almost uniform, but also works
surprisingly well when the distribution of electrons is
strongly inhomogeneous, such as at surfaces and in molecules. The analogous approximation for spin-polarized
systems, which is known as the local spin-density approximation (LSDA), has also proved very successful.
The LDA and LSDA were used throughout the development of density-functional theory and are still widely
used in studies of solids.
In cases when the LDA is not accurate enough, it
seems sensible to try approximating E xc 关 n(r) 兴 in terms
38
Foulkes et al.: Quantum Monte Carlo simulations of solids
of the local density and its gradient. It turns out that
simple gradient expansions are rather badly behaved,
but that better approximations can be devised by expressing ⑀ xc as a carefully chosen nonlinear function of
n(r) and 兩 ⵜn(r) 兩 . This is the idea behind the generalized gradient approximation (Langreth and Mehl, 1983;
Becke, 1988; Perdew et al., 1992), which is now used in
most quantum-chemical applications of densityfunctional theory. Several nonlocal functionals have also
been proposed, of which the best known are the
average-density approximation and weighted-density
approximation of Gunnarsson, Jonson, and Lundqvist
(1979).
Density-functional results obtained using the best
available approximate exchange-correlation energy
functionals are typically an order of magnitude less accurate than good QMC results, but since densityfunctional theory is much less computationally demanding than QMC it has a much wider variety of interesting
applications. Computational experiments based on
density-functional theory are already used in physics,
chemistry, biochemistry, materials science, and geophysics, and are beginning to be used in industrially important fields such as drug design.
E. Quantum Monte Carlo methods
There are many different QMC methods, but this review concentrates on only two: variational quantum
Monte Carlo (VMC) and fixed-node diffusion quantum
Monte Carlo (DMC). Like all QMC methods, these are
closely related to Monte Carlo methods used in classical
statistical mechanics (Nightingale, 1999). In the VMC
method expectation values are calculated via Monte
Carlo integration over the 3N-dimensional space of
electron coordinates. The more sophisticated DMC
method is a projector approach in which a stochastic
imaginary-time evolution is used to improve a starting
trial wave function.
Other Monte Carlo methods, such as auxiliary-field
QMC (reviewed by Senatore and March, 1994) and
path-integral QMC (reviewed by Ceperley, 1995a,
1995b), may also be used to study interacting manyelectron systems. Recent progress along these lines includes the development of the shifted-contour auxiliaryfield method (Baer, Head-Gordon, and Neuhauser,
1998; Baer and Neuhauser, 2000), which has already
been applied to small molecules, and the introduction of
the reptation QMC method (Baroni and Moroni, 1999),
which is an interesting hybrid of path-integral and diffusion QMC. The more established restricted path-integral
fermion Monte Carlo method (Ceperley, 1995b) has
been used to investigate the formation of a gas of molecular hydrogen from a neutral system of electrons and
protons at high temperature (Magro et al., 1996) and to
calculate the forces between protons in an electron gas
(Zong and Ceperley, 1998). Unlike VMC and DMC,
which are zero-temperature methods in which one considers a single wave function, path-integral and
auxiliary-field QMC may both be used to compute exRev. Mod. Phys., Vol. 73, No. 1, January 2001
pectation values at finite temperature. These methods
have not yet been applied to real solids, however, and
we shall not discuss them further in this review. We also
restrict ourselves to continuum models and choose not
to consider lattice models such as the Hubbard model.
Although the methods described in this article may be
applied to lattice problems, other QMC methods such as
the auxiliary-field approach are also useful in such systems (Blankenbecler, Scalapino, and Sugar, 1981).
McMillan (1965) was the first to use the VMC method
in his study of liquid 4He. One of the first applications to
a many-fermion system was by Ceperley, Chester, and
Kalos (1977). A very early calculation by Donsker and
Kac (1950) employed a type of projector Monte Carlo
method, but their algorithm was inefficient and fails for
unbounded potentials such as the Coulomb interaction.
Kalos (1962, 1967) developed the projector Green’sfunction Monte Carlo method and enormously improved its efficiency by introducing the concept of importance sampling, in which a trial or guiding wave
function is used to steer the Monte Carlo moves (Kalos,
Levesque, and Verlet, 1974). The DMC algorithm
(Grimm and Storer, 1971; Anderson, 1975, 1976) discussed here may be viewed as an accurate and convenient approximation to the full Green’s-function Monte
Carlo algorithm and was developed later. Importance
sampling was introduced into the DMC algorithm by
Grimm and Storer (1971).
Because of the fermionic nature of the many-electron
wave function, it must have positive and negative regions. This simple fact underlies the infamous fermion
sign problem, which plagues all projector QMC methods. The search for exact solutions to the sign problem is
still active, but to date no entirely satisfactory exact solution exists. Diffusion Monte Carlo simulations of large
systems therefore use the fixed-node approximation of
Anderson (1975). A many-electron trial wave function is
chosen and used to define a trial nodal surface. [The
nodal surface of a wave function ⌿(r1 ,r2 ,...,rN ) is the
(3N⫺1)-dimensional surface on which ⌿⫽0 and across
which it changes sign.] The fixed-node DMC algorithm
then projects out the many-electron wave function with
the lowest possible energy expectation value consistent
with that fixed nodal surface. The results are not exact
unless the trial nodal surface is exact, but the fixed-node
method is computationally stable and the calculated energy is variational and often very accurate.
The importance-sampled fixed-node fermion DMC algorithm was first applied to the electron gas by Ceperley
and Alder (1980). The first application to solid hydrogen
was in 1987 (Ceperley and Alder, 1987), and the first
VMC and DMC calculations of solids containing heavier
atoms followed soon after (Fahy, Wang, and Louie,
1988, 1990a; Li, Ceperley, and Martin, 1991). Important
recent technical developments have included the introduction of variance minimization techniques to optimize
trial wave functions (Umrigar, Wilson, and Wilkins,
1988) and the use of nonlocal pseudopotentials in VMC
and DMC calculations (Hammond, Reynolds, and
Lester, Jr., 1987; Hurley and Christiansen, 1987; Fahy,
Foulkes et al.: Quantum Monte Carlo simulations of solids
Wang, and Louie, 1988; Mitas, Shirley, and Ceperley,
1991). These improvements have stimulated applications
to a much wider range of systems.
The computer time required to calculate the energy of
a system to some given accuracy using the fermion VMC
and DMC methods effectively scales as N 3 (see Sec.
X.E). Of course QMC does not offer a free lunch—the
price to be paid for this advantageous scaling with N is
that the answer is obtained with a statistical error bar
that decays only as the inverse of the square root of the
computer time. The total amount of time required for
accurate QMC calculations is therefore quite large.
However, QMC algorithms are intrinsically parallel,
which is becoming a significant computational advantage
as the availability of powerful and relatively inexpensive
parallel machines improves.
For small systems the DMC method is capable of
reaching chemical accuracy, which is usually defined as
⬃1 kcal per mole (⬇0.04 eV per molecule). The advantageous scaling with system size means that the attainable accuracy does not fall off rapidly as the number of
electrons N increases. Accurate trial wave functions are
necessary to achieve high accuracy at a reasonable computational cost, but since the DMC wave function is generated stochastically the results are largely free of the
errors caused by the limited basis set used in most other
techniques. By contrast, the quality of results obtained
using the less accurate VMC method is entirely determined by the quality of the trial wave function.
The VMC and DMC methods are best suited to calculating energies because these have the very advantageous zero-variance property; as the trial wave function
approaches the exact ground state (or any other exact
energy eigenstate) the statistical fluctuations in the energy reduce to zero. Other ground-state expectation values, such as static correlation functions, can also be calculated, but the absence of a zero-variance property
makes this more problematic. The VMC and DMC
methods are less well adapted to studying excited states,
but have nevertheless been used successfully to calculate
a wide range of excited-state properties of atoms, molecules, and solids.
III. MONTE CARLO METHODS
A. Statistical foundations
The archetypal example of a Monte Carlo simulation
is the evaluation of a multidimensional integral by sampling the integrand statistically and averaging the
sampled values. In conventional quadrature methods the
accuracy depends on the fineness of the integration
mesh. If one uses a d-dimensional cubic mesh to evaluate a d-dimensional integral using Simpson’s rule, for
example, the error scales as M ⫺4/d , where M is the total
number of mesh points. Therefore, as the dimension d
increases (we have d as large as a few thousand in some
of our applications), the error falls off increasingly
slowly with M. Using Monte Carlo methods, on the
other hand, the statistical error in the value of the inteRev. Mod. Phys., Vol. 73, No. 1, January 2001
39
gral decreases as the square root of the number of sampling points used, regardless of dimensionality. This is a
consequence of the central limit theorem, one of the cornerstones of the mathematical theory of statistics.
Suppose we define a 3N-dimensional vector R by
R⫽ 共 r1 ,r2 ,...,rN 兲 ,
(3.1)
where ri is the position of the ith electron. A particular
value of R is sometimes called a walker, a configuration,
or a psip in the QMC literature. (Note that here and
throughout the rest of Sec. III we consider a system of
spinless fermions for simplicity; Sec. IV discusses the minor extensions required to treat the spin properly.) The
probability density of finding the electrons in the configuration R will be denoted by P(R), where
P共 R兲 ⭓0,
(3.2)
冕
(3.3)
dR P共 R兲 ⫽1.
Let 兵 Rm :m⫽1,M 其 be a set of mutually independent (uncorrelated) configurations distributed according to the
probability distribution P(R). We define a new random
variable Zf by
Zf ⫽
关 f 共 R1 兲 ⫹f 共 R2 兲 ⫹•••⫹f 共 RM 兲兴
,
M
(3.4)
where f(R) is any reasonable function with mean ␮ f and
variance ␴ f2 given by
␮ f⫽
␴ f2 ⫽
冕
冕
dR f 共 R兲 P共 R兲 ,
(3.5)
dR关 f 共 R兲 ⫺ ␮ f 兴 2 P共 R兲 .
(3.6)
Then, under rather general conditions (Feller, 1968), the
central limit theorem tells us that, for large enough M,
Zf is normally distributed with mean ␮ f and standard
deviation ␴ f / 冑M. This implies that, regardless of P(R),
the mean value of a large number of measurements of
some function of R will be a good estimator of the mean
of that function with respect to P(R). Moreover, the
standard deviation of the mean of the measurements
will decrease as 1/冑M, irrespective of the dimension of
the integral.
These ideas can be applied to the evaluation of integrals such as
I⫽
冕
dR g 共 R兲 .
(3.7)
First we introduce an ‘‘importance function’’ P(R) and
rewrite the integral in the form
I⫽
冕
dR f 共 R兲 P共 R兲 ,
(3.8)
where f(R)⬅g(R)/P(R). The importance function is
chosen such that it obeys Eqs. (3.2) and (3.3) and hence
may be interpreted as a probability density. In principle,
the value of I may now be obtained by drawing an infi-
40
Foulkes et al.: Quantum Monte Carlo simulations of solids
nite set of random vectors from the distribution P(R)
and computing the sample average:
再
冎
M
1
f 共 Rm 兲 .
I⫽ lim
M
m⫽1
M→⬁
兺
(3.9)
A Monte Carlo estimate of I may be obtained by averaging over a large but finite sample of vectors drawn
from P(R):
M
I⬇
1
f 共 Rm 兲 .
M m⫽1
兺
(3.10)
The variance of the estimate of the integral is then
␴ f2 /M, which itself can be estimated as
兺冋
M
册
M
2
␴ f2
1
1
⬇
f 共 Rm 兲 ⫺
f 共 Rn 兲 , (3.11)
M M 共 M⫺1 兲 m⫽1
M n⫽1
兺
and an estimate of the size of the error bar on the computed value of I is ⫾ ␴ f / 冑M.
A judicious choice of the importance function P(R)
significantly reduces the variance for a fixed sample size.
It can be readily shown that the importance function
that gives the smallest variance is P(R)⫽ 兩 g(R)/I 兩 .
However, since I is not known, the integrand f(R)
⫽g(R)/P(R) corresponding to this importance function
is also unknown. In general, the best that one can do is
make P(R) as similar to 兩 g(R)/I 兩 as possible.
When using the Monte Carlo method described in
Sec. III.A to evaluate multidimensional integrals, it is
necessary to sample complicated probability distributions in high-dimensional spaces. In general, the normalizations of these distributions are unknown and they are
so complicated that they cannot be sampled directly.
The Metropolis rejection algorithm (Metropolis et al.,
1953) has the great advantage that it allows an arbitrarily complex distribution to be sampled in a straightforward way without knowledge of its normalization.
The Metropolis algorithm generates the sequence of
sampling points Rm by moving a single walker according
to the following rules:
(1) Start the walker at a random position R.
(2) Make a trial move to a new position R⬘ chosen
from some probability density function T(R⬘ ←R). After the trial move the probability that the walker initially
at R is now in the volume element dR⬘ is dR⬘ ⫻T(R⬘
←R).
(3) Accept the trial move to R⬘ with probability
冉
冊
T 共 R←R⬘ 兲 P共 R⬘ 兲
.
T 共 R⬘ ←R兲 P共 R兲
(3.13)
This must be balanced by the number moving from dR⬘
to dR,
⫽A 共 R←R⬘ 兲 T 共 R←R⬘ 兲 n 共 R⬘ 兲 dR⬘ dR,
(3.14)
and hence the equilibrium distribution n(R) satisfies
n 共 R兲 A 共 R←R⬘ 兲 T 共 R←R⬘ 兲
⫽
.
n 共 R⬘ 兲 A 共 R⬘ ←R兲 T 共 R⬘ ←R兲
(3.15)
Since the ratio of Metropolis algorithm acceptance probabilities is
A 共 R←R⬘ 兲 T 共 R⬘ ←R兲 P共 R兲
⫽
,
A 共 R⬘ ←R兲 T 共 R←R⬘ 兲 P共 R⬘ 兲
(3.16)
it follows that
P共 R兲
n 共 R兲
⫽
.
n 共 R⬘ 兲 P共 R⬘ 兲
(3.17)
The equilibrium walker density n(R) is therefore proportional to P(R), and the probability of finding any
given walker in dR is P(R)dR as required. A rigorous
derivation of this result is given by Feller (1968).
C. Variational Monte Carlo
(3.12)
If the trial move is accepted, the point R⬘ becomes the
next point on the walk; if the trial move is rejected, the
point R becomes the next point on the walk. If P(R) is
high, most trial moves away from R will be rejected and
the point R may occur many times in the set of points
making up the random walk.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
dR⬘ A 共 R⬘ ←R兲 T 共 R⬘ ←R兲 ⫻n 共 R兲 dR.
A 共 R⬘ ←R兲 T 共 R⬘ ←R兲 n 共 R兲 dR dR⬘
B. The Metropolis algorithm
A 共 R⬘ ←R兲 ⫽Min 1,
(4) Return to step 2 and repeat.
The initial points generated by this algorithm depend
on the starting point and should be discarded. Eventually, however, the simulation settles down and sets of
points snipped out of the random walk are distributed
according to P(R).
To understand how this algorithm works, imagine an
enormous number of walkers all executing random
walks according to the above rules. After an equilibration period, we assume that the distribution of walkers
settles down to a unique steady equilibrium state in
which the average number of walkers in the volume element dR is denoted by n(R)dR. Once the equilibrium
has been established, we also assume that the average
number of walkers moving from dR to dR⬘ in one time
step is the same as the number moving from dR⬘ to dR.
Our everyday experience of diffusion and other randomwalk processes suggests that these assumptions are probably correct, but readers who wish to see a clear derivation of the conditions under which they hold are
referred to Feller (1968).
Since the probability that the next move of a walker at
R takes it to dR⬘ is dR⬘ A(R⬘ ←R)T(R⬘ ←R), the average number moving from dR to dR⬘ during a single
move will be
The variational Monte Carlo (VMC) method is the
simpler of the two quantum Monte Carlo methods discussed in this review. It is based on a combination of the
variational principle and the Monte Carlo evaluation of
integrals discussed in Sec. III.A. The VMC method relies on the availability of a trial wave function ⌿ T that is
a reasonably good approximation of the true ground-
41
Foulkes et al.: Quantum Monte Carlo simulations of solids
state wave function. The subject of how to produce good
trial wave functions is dealt with in Sec. IV. The trial
wave function must satisfy some basic conditions. Both
⌿ T and ⵜ⌿ T must be continuous wherever the potential
is finite, and the integrals 兰 ⌿ T* ⌿ T and 兰 ⌿ T* Ĥ⌿ T must
exist. To keep the variance of the energy finite we also
require 兰 ⌿ T* Ĥ 2 ⌿ T to exist.
The expectation value of Ĥ evaluated with a trial
wave function ⌿ T provides a rigorous upper bound on
the exact ground-state energy E 0 :
E V⫽
冕
冕
⌿ T* 共 R兲 Ĥ⌿ T 共 R兲 dR
⭓E 0 .
(3.18)
⌿ T* 共 R兲 ⌿ T 共 R兲 dR
In a VMC simulation this bound is calculated using the
Metropolis Monte Carlo method. Equation (3.18) is rewritten in the form
E V⫽
冕
FIG. 1. Flow chart illustrating the variational Monte Carlo
(VMC) algorithm.
兩 ⌿ T 共 R兲 兩 2 关 ⌿ T 共 R兲 ⫺1 Ĥ⌿ T 共 R兲兴 dR
冕
,
(3.19)
兩 ⌿ T 共 R兲 兩 2 dR
and the Metropolis algorithm is used to sample a set of
points 兵 Rm :m⫽1,M 其 from the configuration-space
probability density P(R)⫽ 兩 ⌿ T (R) 兩 2 / 兰兩 ⌿ T (R) 兩 2 dR. At
each of these points the ‘‘local energy’’ E L (R)
⫽⌿ T (R) ⫺1 Ĥ⌿ T (R) is evaluated and the average energy accumulated:
M
1
E V⬇
E 共 R 兲.
M m⫽1 L m
兺
(3.20)
The trial moves are usually sampled from a Gaussian
centered on the current position of the walker, the variance of the Gaussian being chosen such that average
acceptance probability is roughly 50% or such that the
diffusion constant of the random walk is maximized. The
flow chart shown in Fig. 1 illustrates how a typical VMC
simulation works.
Expectation values of operators other than the Hamiltonian may also be expressed as 3N-dimensional integrals and evaluated in an analogous way. The variance
of the QMC result for any given observable depends on
the details of the integrand, suggesting that it may be
possible to reduce the variance by modifying the estimator in such a way that its expectation value is unaffected.
Very recently, Assaraf and Caffarel (1999) have demonstrated that a significant variance reduction may be obtained in both VMC and DMC using this technique, although their approach has been tested only in small
systems so far.
D. Diffusion Monte Carlo
1. The imaginary-time Schrödinger equation
Diffusion Monte Carlo (DMC) is a stochastic projector method for solving the imaginary-time many-body
Schrödinger equation,
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
⫺ ⳵ t ⌽ 共 R,t 兲 ⫽ 共 Ĥ⫺E T 兲 ⌽ 共 R,t 兲 ,
(3.21)
where t is a real variable measuring the progress in
imaginary time, R⫽(r1 ,r2 , . . . ,rN ) is a 3N-dimensional
vector specifying the coordinates of all N electrons, and
E T is an energy offset. Equation (3.21) may be rewritten
in the integral form
冕
G 共 R←R⬘ , ␶ 兲 ⌽ 共 R⬘ ,t 兲 dR⬘ ,
(3.22)
G 共 R←R⬘ , ␶ 兲 ⫽ 具 R兩 exp关 ⫺ ␶ 共 Ĥ⫺E T 兲兴 兩 R⬘ 典
(3.23)
⌽ 共 R,t⫹ ␶ 兲 ⫽
where
is a Green’s function that obeys the same equation as
the wave function,
⫺ ⳵ t G 共 R←R⬘ ,t 兲 ⫽„Ĥ 共 R兲 ⫺E T …G 共 R←R⬘ ,t 兲 , (3.24)
with the initial condition G(R←R⬘ ,0)⫽ ␦ (R⫺R⬘ ). Using the spectral expansion
exp共 ⫺ ␶ Ĥ 兲 ⫽
兺i 兩 ⌿ i 典 exp共 ⫺ ␶ E i 兲 具 ⌿ i兩 ,
(3.25)
one can express the Green’s function as
G 共 R←R⬘ , ␶ 兲 ⫽
兺i ⌿ i共 R兲 e ⫺ ␶ (E ⫺E
i
T)
⌿*
i 共 R⬘ 兲 , (3.26)
where 兵 ⌿ i 其 and 兵 E i 其 denote the complete sets of eigenfunctions and eigenvalues of Ĥ, respectively. Equations
(3.21) and (3.22) are equivalent, as can be readily demonstrated by differentiating Eq. (3.22) with respect to ␶
and setting ␶ ⫽0.
It is straightforward to show that as ␶ →⬁ the operator
exp关⫺␶(Ĥ⫺ET)兴 projects out the lowest eigenstate 兩 ⌿ 0 典
that has nonzero overlap with the chosen 兩 ⌽(t⫽0) 典
⫽ 兩 ⌽ init典 :
42
Foulkes et al.: Quantum Monte Carlo simulations of solids
lim 具 R兩 exp关 ⫺ ␶ 共 Ĥ⫺E T 兲兴 兩 ⌽ init典
␶ →⬁
⫽ lim
冕
⫽ lim
兺 ⌿ i共 R兲 exp关 ⫺ ␶ 共 E i ⫺E T 兲兴 具 ⌿ i兩 ⌽ init典
␶ →⬁
G 共 R←R⬘ , ␶ 兲 ⌽ init共 R⬘ 兲 dR⬘
␶ →⬁ i
⫽ lim ⌿ 0 共 R兲 exp关 ⫺ ␶ 共 E 0 ⫺E T 兲兴 具 ⌿ 0 兩 ⌽ init典 .
ˆ
(3.27)
By adjusting E T to equal E 0 , one can make the exponential factor in the last line of Eq. (3.27) constant, while
the higher states in the previous line are all exponentially damped because their energies are higher than E 0 .
This fundamental property of the projector exp关⫺␶(Ĥ
⫺ET)兴 is the basis of the diffusion Monte Carlo method
and similar projector-based approaches.
Let us for a moment neglect the potential terms in the
Hamiltonian so that Eq. (3.21) simplifies to
N
兺
(3.28)
This is a diffusion equation in a 3N-dimensional space,
the Green’s function for which is a 3N-dimensional
Gaussian with variance ␶ in each dimension,
冋
G 共 R←R⬘ , ␶ 兲 ⫽ 共 2 ␲ ␶ 兲 ⫺3N/2 exp
册
⫺ 兩 R⫺R⬘ 兩 2
.
2␶
discrete
兺k ␦ 共 R⫺Rk 兲 .
(3.30)
The Green’s function of Eq. (3.29) can be interpreted as
a transition probability density for the evolution of the
walkers. By substituting Eq. (3.30) into Eq. (3.22) we
obtain
discrete
⌽ 共 R,t⫹ ␶ 兲 哫
兺k G 共 R←Rk , ␶ 兲 .
(3.31)
The result is a set of Gaussians each with a variance of
3N ␶ . In order to restore the original discrete representation, one samples each Gaussian by a new delta function that defines the evolved position of the walker. The
procedure of propagation/resampling is then repeated
until convergence is reached. The result of this process is
a set of samples distributed according to the solution of
Eq. (3.28).
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
ˆ
ˆ
(3.32)
If we write Ĥ⫽T̂⫹V̂, where T̂ is the full N-electron
kinetic-energy operator and V̂ is the total potentialenergy operator (including both electron-ion and
electron-electron terms), application of the TrotterSuzuki formula with Â⫽T̂ and B̂⫽V̂⫺E T gives
ˆ
G 共 R←R⬘ , ␶ 兲 ⫽ 具 R兩 e ⫺ ␶ (T ⫹V̂⫺E T ) 兩 R⬘ 典
ˆ
⬇e ⫺ ␶ [V(R)⫺E T ]/2具 R兩 e ⫺ ␶ T 兩 R⬘ 典
⫻e ⫺ ␶ [V(R⬘ )⫺E T ]/2.
(3.33)
The approximate Green’s function for small ␶ is therefore given by (Reynolds et al., 1982)
冋
G 共 R←R⬘ , ␶ 兲 ⬇ 共 2 ␲ ␶ 兲 ⫺3N/2 exp ⫺
共 R⫺R⬘ 兲 2
2␶
册
⫻exp关 ⫺ ␶ 关 V 共 R兲 ⫹V 共 R⬘ 兲 ⫺2E T 兴 /2兴 ,
(3.29)
It is easy to verify that this Gaussian form tends to
␦ (R⫺R⬘ ) as ␶ →0 and that it satisfies Eq. (3.28).
In the theory of stochastic processes Eq. (3.28) is
called the master equation of a diffusion stochastic process (Karlin and Taylor, 1981) and its solution ⌽(R,t)
describes the distribution of the diffusing Brownian particles in space and time. The correspondence Brownian
particles ↔ distribution function works both ways, and
we can represent the distribution ⌽(R,t) by a set of
discrete Brownian sampling points or random walkers:
⌽ 共 R,t 兲 哫
ˆ
e ⫺ ␶ (A ⫹B̂) ⫽e ⫺ ␶ B /2e ⫺ ␶ A e ⫺ ␶ B /2⫹O关 ␶ 3 兴 .
␶ →⬁
1
⳵ t ⌽ 共 R,t 兲 ⫽
ⵜ 2 ⌽ 共 R,t 兲 .
2 i⫽1 i
Now consider the full Hamiltonian, which includes
both kinetic and potential-energy terms. Except in a few
special cases, once the particles are interacting we do
not know explicit expressions for the exact Green’s
function and have to use approximations. An approximation to the Green’s function can be obtained using
the Trotter-Suzuki formula for the exponential of a sum
of operators  and B̂:
where the error is proportional to ␶ 3 .
The factor
P⫽exp关 ⫺ ␶ „V 共 R兲 ⫹V 共 R⬘ 兲 ⫺2E T …/2兴
(3.34)
(3.35)
acts as a time-dependent renormalization (reweighting)
of the diffusion Green’s function. This change of normalization can be incorporated into the process of
walker evolution in various ways. One possibility is to
assign each walker a weight and accumulate the product
of weights during the propagation. However, this option
does not give the most efficient sampling since the
weights of the walkers rapidly become very different
and, in the long time limit, one dominates exponentially
over the rest. A better alternative is to use the branching
or birth/death algorithm, in which P determines the
number of walkers that survive to the next step (Reynolds et al., 1982). In its simplest form the procedure is
as follows:
(1) If P⬍1 the walker continues its evolution with
probability P.
(2) If P⭓1 the walker continues; in addition, at the
same position, a new walker is created with probability
P⫺1.
Both possibilities can be conveniently coded as a
single command,
M new⫽INT共 P⫹ ␩ 兲 ,
(3.36)
where M new is the number of walkers evolving to the
next step at a given position, INT denotes the integer
part of a real number, and ␩ is a random number drawn
Foulkes et al.: Quantum Monte Carlo simulations of solids
43
cessfully applied to small molecules, but practical applications are restricted to few-electron systems (Diedrich
and Anderson, 1992). Nevertheless, for small enough
systems, very impressive results have been obtained. For
example, the energy surface for the reaction H2⫹H has
been evaluated at a few thousand points with an astonishing accuracy of 0.01 kcal/mol (Wu, Kuppermann, and
Anderson, 1999), improving existing results by an order
of magnitude. The investigation of approximation-free
algorithms for fermions is a very active area of current
research, but deeper analysis is outside the scope of this
review; we refer the reader to the literature (Hammond,
Lester, and Reynolds, 1994; Kalos and Pederiva, 1999).
2. The fixed-node approximation
FIG. 2. Illustration of the walker evolution in the diffusion
Monte Carlo (DMC) method. The example shows a onedimensional problem in which a single particle is confined by a
potential well V(x). The initial walker distribution samples a
uniform ⌿ init . As the imaginary-time propagation proceeds,
the distribution converges towards the ground state ⌿ 0 . Note
the occasional disappearance of walkers in the region of high
potential energy and the proliferation around the potential
minimum.
from a uniform distribution on the interval 关 0,1兴 . From
the expression for P it is clear that in regions of high
potential energy the walkers disappear, while in regions
of low potential energy they proliferate. The branching
algorithm therefore transforms the weight accumulation
in the low-energy regions into an increase in the density
of walkers there.
The energy offset E T , which determines the overall
asymptotic renormalization [see Eq. (3.27)], is used to
control the total population of walkers. During the
propagation E T is occasionally adjusted so that the overall number of walkers fluctuates around a desired mean
value. Typical mean values of the number of walkers
used in simulations are between 102 and 103 . Figure 2
illustrates this DMC algorithm for the case of a single
particle moving in a one-dimensional potential well.
So far we have assumed that the wave function is positive everywhere. However, the fermion antisymmetry
implies that many-fermion wave functions cannot be
positive everywhere and must take both positive and
negative values. Unfortunately, probabilistic methods
such as DMC can handle only positive distributions. A
straightforward generalization of the DMC algorithm,
such as an assignment of sign variables to walkers, while
formally correct, leads to the fermion sign problem and
an exponentially decaying signal-to-noise ratio. More
elaborate algorithms with walker signs have been sucRev. Mod. Phys., Vol. 73, No. 1, January 2001
Fixed-node DMC (Anderson, 1975, 1976; Moskowitz
et al., 1982; Reynolds et al., 1982) is an alternative
method for dealing with the fermion antisymmetry. Although not exact, it gives ground-state energies that satisfy a variational principle and are usually very accurate.
Furthermore, unlike the exact methods described above,
the fixed-node algorithm is stable in large systems. The
fixed-node approximation is therefore used in almost all
current large-scale applications of DMC. The version
described here assumes that the ground-state wave function is real and hence works only in systems with timereversal symmetry (i.e., a real Hamiltonian). There is,
however, a successful generalization known as the fixedphase approximation (Ortiz, Ceperley, and Martin,
1993) for use in systems without time-reversal symmetry. This is particularly useful for studying interacting
electrons in an applied magnetic field or states with nonzero angular momentum.
The basic idea behind fixed-node DMC is very simple.
A trial many-electron wave function is chosen and used
to define a trial many-electron nodal surface. In a threedimensional system containing N electrons, the manyelectron trial state is a function of 3N variables (the x,
y, and z coordinates of each electron) and the trial
nodal surface is the (3N⫺1)-dimensional surface on
which that function is zero and across which it changes
sign. The fixed-node DMC algorithm then produces the
lowest-energy many-electron state with the given nodal
surface. Fixed-node DMC may therefore be regarded as
a variational method that gives exact results if the trial
nodal surface is exact. It differs from VMC, however, in
that no assumptions are made about the functional form
of the state between the nodes.
a. One-electron example
The fixed-node approximation is best introduced by
means of a simple example. Figure 3 illustrates the standard DMC algorithm for the case of a particle in a box.
Walkers initially distributed according to the starting
state (the broken line) diffuse randomly until they cross
one of the box walls, at which point they are removed
from the simulation. This boundary condition follows
from Eq. (3.35), which shows that the reweighting factor
P is zero where the potential is infinite.
44
Foulkes et al.: Quantum Monte Carlo simulations of solids
FIG. 3. A possible initial walker density distribution (dashed
line) and the final walker distribution (solid line) for a DMC
simulation of a particle in a box.
The absorption of walkers at the walls guarantees that
the average walker density tends to zero at the edges of
the box and hence that the eigenstate obtained satisfies
the correct boundary conditions. The energy offset E T is
chosen such that P is slightly greater than unity within
the box; this ensures that the diffusing walkers multiply
steadily and counteracts the steady loss of walkers at the
walls. Regardless of the shape of the initial distribution,
the walker density settles down to the final state shown,
which is proportional to the ground-state wave function.
How might this algorithm be adapted to give the first
excited state, which has odd parity and a node at the box
center? One approach is to start with a trial function of
odd parity, such as that shown in Fig. 4(a), and write it
as the difference between the two non-negative distributions shown in Fig. 4(b):
⌽ 共 x,t⫽0 兲 ⫽⌽ ⫹ 共 x,t⫽0 兲 ⫺⌽ ⫺ 共 x,t⫽0 兲 ,
(3.37)
where
1
⌽ ⫹ 共 x,t⫽0 兲 ⫽ 关 兩 ⌽ 共 x,t⫽0 兲 兩 ⫹⌽ 共 x,t⫽0 兲兴 ,
2
(3.38)
1
⌽ ⫺ 共 x,t⫽0 兲 ⫽ 关 兩 ⌽ 共 x,t⫽0 兲 兩 ⫺⌽ 共 x,t⫽0 兲兴 .
2
(3.39)
The Hamiltonian commutes with the parity operator
and so preserves the parity of the starting state
⌽(x,t⫽0). Furthermore, since the imaginary-time
FIG. 4. Decomposition of an antisymmetric starting state into
non-negative components. Panel (a) shows a possible starting
state ⌽(x,t⫽0) for a DMC simulation of the lowest state with
odd parity of a particle in a box; (b) shows the two nonnegative distributions ⌽ ⫹ (x,t⫽0) and ⌽ ⫺ (x,t⫽0) into which
⌽(x,t⫽0) is decomposed in Eq. (3.37).
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Schrödinger equation is linear, the two positive distributions ⌽ ⫹ (x,t) and ⌽ ⫺ (x,t) may be obtained using two
separate DMC simulations. The odd function ⌽(x,t)
may then be calculated by subtracting ⌽ ⫺ (x,t) from
⌽ ⫹ (x,t). This is the simplest of the exact solutions to
the sign problem mentioned above, but it is clear that
both ⌽ ⫹ and ⌽ ⫺ tend to the even-parity ground state as
t→⬁. The odd-parity components of interest decay like
exp(⫺⌬Et) relative to the even components, where ⌬E
is the difference between the odd- and even-parity
ground-state energies. In practice, therefore, the statistical noise in ⌽ ⫹ and ⌽ ⫺ soon swamps the exponentially
decaying odd-parity signal and the subtraction gives
nonsense.
The fixed-node solution to this problem works as follows. Imagine introducing an extra absorbing barrier at
the center of the box, dividing the interior into two separate simulation regions. The DMC walkers are initially
scattered throughout both regions, but the simulations
on the left and right then progress independently. In the
large-t limit, the walker densities in the two regions are
proportional to the lowest-energy eigenfunctions satisfying zero boundary conditions at both the box walls and
the absorbing barrier. The odd-parity eigenfunction of
interest satisfies these same Dirichlet boundary conditions, which determine the eigenfunctions within each
region. The t→⬁ walker densities within each region
must therefore be proportional to the odd-parity eigenfunction in that region.
b. Many-electron version
Although the nodes of the many-electron ground state
of a solid are very complicated, the fixed-node approximation works in the same way. If the nodes are known
exactly, absorbing barriers may be placed everywhere
on the nodal surface, dividing up the configuration space
into a set of disjoint nodal pockets. Parallel DMC simulations are then carried out in all the nodal pockets, and
the solution within each pocket tends to ⫾ the exact
ground-state wave function in that pocket.
The wave function ⌿(x 1 ,x 2 ,...,x N ) of N spinless
electrons in one dimension must be zero whenever any
two electrons coincide. These coincidence conditions define (N⫺1)-dimensional coincidence planes passing
through the N-dimensional configuration space. Given
that the nodal surface is also (N⫺1) dimensional, it is
possible that the coincidence planes exhaust the nodal
surface and hence determine it exactly. Ceperley (1991)
has shown that this is indeed the case for the ground
state of a one-dimensional system. In d-dimensional systems, however, the coincidence planes are (dN⫺d) dimensional, whereas the nodal surface is (dN⫺1) dimensional. For d⬎1 the coincidence planes are therefore no
more than a framework through which the nodal surface
must pass and cannot determine it completely.
The fixed-node approximation (Anderson, 1975, 1976)
circumvents this difficulty by using an approximate
nodal surface, which is normally obtained from a variational trial wave function. We show below that fixed-
Foulkes et al.: Quantum Monte Carlo simulations of solids
45
of the lowest-energy nodeless wave function within the
most favorable pockets. This selection of favorable
pockets does indeed happen in some excited-state DMC
simulations, but Ceperley (1991) proved that the nodal
pockets of the ground state of any Hamiltonian with a
reasonable local potential are all equivalent by symmetry. As long as the trial nodal surface is that of the
ground state of some local Hamiltonian (perhaps the
LDA one), this tiling theorem ensures that all nodal
pockets are equally favorable. A derivation of the tiling
theorem is given below.
c. The fixed-node variational principle
FIG. 5. A two-dimensional slice through the 321-dimensional
nodal surface of a two-dimensional electron gas containing 161
spin-up electrons subject to periodic boundary conditions
(Ceperley, 1991). The slice was defined by freezing 160 of the
electrons at the random positions shown by the open circles
and leaving the one remaining electron, represented by the
filled circle, free to move around.
node DMC energies are variational, i.e., the fixed-node
DMC energy is always greater than or equal to the exact
ground-state energy, with the equality holding only
when the trial nodal surface is exact. Since the DMC
energy is normally a smooth function of the trial nodal
surface, the errors in the energy are normally second
order in the errors in the trial nodal surface. In practice,
the errors in fixed-node DMC energies are typically
about 5% of the correlation energy. Lest you imagine
that this high accuracy arises because nodal surfaces are
in some sense simple or easy to guess, look at Fig. 5
(which is taken from Ceperley, 1991). The little that is
understood about real nodal surfaces is discussed by
Caffarel and Claverie (1988), Ceperley (1991), Glauser
et al. (1992), and Foulkes, Hood, and Needs (1999).
The implementation of the fixed-node DMC algorithm is straightforward. One scatters DMC walkers
throughout the configuration space and moves them in
the usual way. The only new ingredient is that after every DMC move the sign of the trial wave function is
checked and the walker deleted if it has crossed the trial
nodal surface. (N.B. In the importance-sampled algorithm discussed in Sec. III.D.3, this deletion step is replaced by a rejection step.) Within each pocket, the
fixed-node DMC algorithm projects out the lowestenergy nodeless wave function satisfying zero boundary
conditions on the enclosing nodal surface.
The only communication between the simulations in
different pockets is via the energy offset E T , which is
gradually adjusted to keep the total walker population
roughly constant. If some pockets are more favorable
than others (that is, if they have a greater average value
of the reweighting factor P), the walker population becomes more and more concentrated in these pockets.
The calculated energy therefore tends to the eigenvalue
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Consider the nodal pockets of an antisymmetric trial
function ⌿ T (R). We can group these pockets into
classes equivalent by permutation symmetry as follows.
Pick a nodal pocket at random and color it blue. Now
pick a point R within the blue pocket and apply a permutation P to it. This maps R to PR, which may or may
not be in the same nodal pocket as R. If the new point
PR is in a new nodal pocket, color that pocket blue as
well. Repeat this process for every permutation P until
all nodal pockets equivalent to the first one by permutation symmetry have been found and colored blue.
There are now two possibilities: either the blue pockets fill up the configuration space, in which case all nodal
pockets are equivalent by permutation symmetry, or
they do not. If they do not, pick one of the remaining
pockets at random, color it red, and repeat the same
procedure, applying permutation operators until all
equivalent pockets have been found and colored red. If
there are still pockets left, pick new colors and repeat
the process until the configuration space is filled by
pockets of different colors. The nodal pockets are thus
divided into equivalence classes, with all the members of
a class related by permutation symmetry.
If ⌿ T (R) is used to define the nodal surface for a
fixed-node DMC simulation, the DMC walkers will
eventually concentrate in the nodal pockets of the
lowest-energy class. Within any such pocket v ␣ , the
walker distribution will tend to the pocket ground state
⌿ ␣0 (R), which is the lowest-energy real normalized
wave function that is zero outside v ␣ and satisfies the
fixed-node boundary conditions on the surface of v ␣ .
This function satisfies the equation
Ĥ⌿ ␣0 共 R兲 ⫽E ␣0 ⌿ 0␣ 共 R兲 ⫹ ␦ ␣ .
␣
(3.40)
Within v the pocket ground state is a nodeless eigenfunction of the Schrödinger equation with eigenvalue
E 0␣ ; outside v ␣ it is zero. Since ⌿ ␣0 (R) normally approaches the surface of v ␣ with nonzero slope, its gradient changes discontinuously as the surface is crossed; the
action of the kinetic-energy operator on these gradient
discontinuities produces the delta functions denoted ␦ ␣ .
The DMC energy for pocket v ␣ (or any symmetryequivalent pocket) is equal to the pocket eigenvalue E ␣0 .
Moskowitz et al. (1982) and Reynolds et al. (1982)
proved that E ␣0 is greater than or equal to the exact
ground-state energy E 0 . Here we follow Reynolds’s
46
Foulkes et al.: Quantum Monte Carlo simulations of solids
proof, which starts from a single-pocket ground state
⌿ ␣0 (R) and uses the permutations P to construct a real
antisymmetric wave function,
⌿̃ ␣0 共 R兲 ⫽
1
NP
兺P 共 ⫺1 兲 ␨ ⌿ 0␣共 PR兲 ⬅Â⌿ ␣0 共 R兲 ,
P
(3.41)
where N P is the total number of permutations and ␨ P is
the parity of permutation P. The pocket ground state
⌿ ␣0 (R) has a nonzero overlap with the antisymmetric
trial function ⌿ T (R) and must therefore have a nonzero
antisymmetric component. The projection operator Â
picks out this component and so ⌿̃ ␣0 (R) cannot be zero.
Since ⌿ ␣0 (R) is zero everywhere on the nodal surface of
⌿ T (R), and since that nodal surface is invariant under
all permutations P, the function ⌿̃ ␣0 (R) is also zero everywhere on the nodal surface of ⌿ T (R).
The real antisymmetric function ⌿̃ ␣0 (R) is now substituted into the standard quantum-mechanical variational
principle to give
E 0⭐
⫽
⫽
E 0⫽
兰 ⌿̃ ␣0 ⌿̃ 0␣ dR
兰 ⌿̃ ␣0 ĤÂ⌿ 0␣ dR
兰 ⌿̃ ␣0 Â⌿ ␣0 dR
兰 ⌿̃ ␣0 Ĥ⌿ 0␣ dR
兰 ⌿̃ ␣0 ⌿ ␣0 dR
具 ⌿̃ 0 兩 ⌿̃ 0 典
.
(3.43)
(3.42)
where we have used the fact that  commutes with Ĥ,
that it is self-adjoint, and that it is idempotent [so
Â⌿̃ ␣0 (R)⫽⌿̃ ␣0 (R)]. The delta-function terms appearing
in Ĥ⌿ ␣0 do not contribute to the energy expectation
value because they occur on the fixed nodal surface
where ⌿̃ 0␣ (R)⫽0. If the nodal surface of ⌿ T (R) is the
same as the nodal surface of the exact ground state, the
equality holds and the fixed-node DMC energy E 0␣ is
equal to E 0 ; but if ⌿ T (R) does not have the correct
nodal surface then E ␣0 ⬎E 0 . This is the fixed-node variational principle. Since the calculated ground-state energy E ␣0 is minimized and equal to the exact groundstate energy when the trial nodal surface is exact, the
error in the calculated energy is in general second order
in the error in the nodal surface.
3. Importance sampling
The simple version of the DMC algorithm outlined in
the previous section is usually spectacularly inefficient,
mainly because the renormalization factor P from Eq.
(3.35) may fluctuate wildly from step to step. In some
circumstances, such as when two charged particles coincide, P can even diverge, making the renormalization
process ill defined.
One can overcome these difficulties by carrying out an
importance-sampling transformation using a ‘‘trial’’ or
‘‘guiding’’ wave function ⌿ T (R) (Grimm and Storer,
1971; Ceperley and Kalos, 1979; Reynolds et al., 1982).
Let us multiply Eq. (3.21) by ⌿ T (R) and introduce a
new function f(R,t)⫽⌽(R,t)⌿ T (R). After rearranging
terms we obtain
1
⫺ ⳵ t f 共 R,t 兲 ⫽⫺ ⵜ 2 f 共 R,t 兲 ⫹ⵜ• 关 vD 共 R兲 f 共 R,t 兲兴
2
d. The tiling theorem
Given any many-electron Hamiltonian with a reasonable local potential, the tiling theorem (Ceperley, 1991)
states that all the ground-state nodal pockets belong to
the same class. (The definition of a class was discussed in
Sec. III.D.2.C above.) This is a many-fermion generalization of the statement that the ground state of a oneparticle Hamiltonian with a local potential has no nodes.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
具 ⌿̃ 0 兩 Ĥ 兩 ⌿̃ 0 典
We have therefore reached a contradiction: only the exact ground state (or a linear combination of degenerate
ground states) could give an energy expectation value
equal to E 0 ; but ⌿̃ 0 (R) cannot be an exact eigenstate
unless the potential diverges almost everywhere on the
nodal surface separating the blue pockets from the rest
of configuration space. If there is a region of this surface
on which the potential is finite, any exact eigenstate
within the blue region is bound to leak out into the surrounding regions. This implies that the exact ground
state cannot have more than one class of nodal pocket.
兰 ⌿̃ ␣0 Ĥ⌿̃ 0␣ dR
⫽E ␣0 ,
The tiling theorem and the fixed-node variational principle are very closely related, as the following derivation
demonstrates.
Assume, for the sake of argument, that ⌿ 0 (R) is the
ground state of a Hamiltonian Ĥ with a reasonable local
potential, but that the nodal pockets of ⌿ 0 (R) fall into
two or more different classes. Color the nodal pockets
belonging to one of these classes blue and consider the
function ⌿̃ 0 (R), defined to equal ⌿ 0 (R) in the blue
pockets and zero elsewhere. This function is properly
antisymmetric but has gradient discontinuities on the
nodal surface separating the blue pockets from the rest
of configuration space. The action of the kinetic-energy
operator turns these gradient discontinuities into delta
functions, but since they occur where ⌿̃ 0 (R)⫽0 they do
not affect the energy expectation value,
⫹ 关 E L 共 R兲 ⫺E T 兴 f 共 R,t 兲 ,
(3.44)
where ⵜ⫽(ⵜ 1 ,ⵜ 2 ,...,ⵜ N ) is the 3N-dimensional gradient operator, vD (R) is the 3N-dimensional drift velocity
defined by
vD 共 R兲 ⫽ⵜ ln 兩 ⌿ T 共 R兲 兩 ⫽⌿ T 共 R兲 ⫺1 ⵜ⌿ T 共 R兲 ,
and
(3.45)
47
Foulkes et al.: Quantum Monte Carlo simulations of solids
E L 共 R兲 ⫽⌿ T 共 R兲 ⫺1 Ĥ⌿ T 共 R兲
(3.46)
is the local energy as used in VMC simulations.
The corresponding integral equation is modified accordingly,
f 共 R,t⫹ ␶ 兲 ⫽
冕
G̃ 共 R←R⬘ , ␶ 兲 f 共 R⬘ ,t 兲 dR⬘ ,
(3.47)
where the modified Green’s function G̃(R←R⬘ , ␶ ) is by
definition equal to ⌿ T (R)G(R←R⬘ , ␶ )⌿ T (R⬘ ) ⫺1 . The
short-time approximation to G̃(R←R⬘ , ␶ ) may be derived using techniques analogous to those employed in
the derivation of Eq. (3.34). The result is
G̃ 共 R←R⬘ , ␶ 兲 ⬇G d 共 R←R⬘ , ␶ 兲 G b 共 R←R⬘ , ␶ 兲 ,
(3.48)
where
G d 共 R←R⬘ , ␶ 兲
冋
⫽ 共 2 ␲ ␶ 兲 ⫺3N/2exp ⫺
关 R⫺R⬘ ⫺ ␶ vD 共 R⬘ 兲兴 2
2␶
册
(3.49)
and
G b 共 R←R⬘ , ␶ 兲
⫽exp兵 ⫺ ␶ 关 E L 共 R兲 ⫹E L 共 R⬘ 兲 ⫺2E T 兴 /2其 .
(3.50)
Importance sampling has several consequences. First,
the density of walkers is enhanced in the regions where
⌿ T is large and vice versa. This is because the drift velocity vD (R) carries the walkers along in the direction of
increasing 兩 ⌿ T 兩 . Second, the exponent in the reweighting term [see Eq. (3.35)] now contains the local energy
E L ⫽⌿ T⫺1 Ĥ⌿ T instead of the potential energy. This is
crucial because for a good trial function the local energy
is close to the ground-state energy eigenvalue and
roughly constant, so the population fluctuations are
much diminished. The overall improvement in efficiency
due to importance sampling can be several orders of
magnitude. Without this improvement, DMC simulations involving hundreds or, in recent attempts, thousands of electrons would not be possible.
The importance-sampling transformation is also extremely helpful in satisfying the fixed-node constraint.
Whenever a walker approaches the nodal surface of
⌿ T , the drift velocity grows and carries it away. In fact,
the drift near the nodal surface is proportional to 1/x,
where x is the distance in the direction normal to that
surface. Furthermore, unless the trial function happens
to be an exact eigenfunction, the local energy also diverges like 1/x. It is possible to show that the divergence
of the drift reduces the probability of crossing the node
to zero, and hence, since we also neglect the effects of
the infinite local energy on the nodal surface itself, that
the importance-sampling transformation automatically
enforces the fixed-node constraint.
In real simulations, however, the Green’s function of
Eq. (3.48) is used. This is not exact, and so walkers on
rare occasions attempt to cross the nodal surface. The
problem can be understood by realizing that the drift
term in Eq. (3.49) is proportional to ␶, while the stanRev. Mod. Phys., Vol. 73, No. 1, January 2001
dard deviation of the Gaussian, which arises from the
diffusion of the walker, is proportional to 冑␶ . In order to
ensure that the short-time approximation is accurate,
most simulations use values of ␶ much smaller than 1
a.u., and so the diffusion can occasionally overcome the
drift and the walker can cross the node. This node crossing reflects a failure of the approximate Green’s function to describe the region close to the node, and of
course becomes less likely as the time step ␶ tends to
zero. In order to fulfill the fixed-node constraint one can
either eliminate any walker that crosses a node (in other
words, apply absorbing boundary conditions as discussed in Sec. III.D.2), or else reject the move and keep
the walker at its original position. The choice of procedure has no effect in the limit of small enough time step
(when the probability of crossing the nodal surface tends
to zero), but it turns out that the rejection algorithm
gives smaller time-step errors.
The Green’s function of Eq. (3.48) is usually a reasonable approximation in regions where the trial wave function is smooth and nonzero. Close to the fixed nodal
surface or to a singularity in the potential, however, this
is not the case, and quantities such as the drift velocity
and local energy may even diverge. The approximate
Green’s function of Eq. (3.48) is then poor and the resulting bias can be significant (Umrigar, Nightingale, and
Runge, 1993).
The simplest remedy is to take smaller time steps, although this makes the calculation rather inefficient. A
better remedy is to improve the Green’s function. A
clever idea due to Ceperley, Kalos, and Lebowitz (1981)
and Reynolds et al. (1982) is to incorporate a rejection
step into the propagation governed by G d (R←R⬘ , ␶ ).
The rejection step is designed to impose a detailed balance condition, which is motivated by two arguments.
First, the exact importance-sampled Green’s function
fulfills the detailed balance condition
G̃ 共 R←R⬘ , ␶ 兲 ⌿ T 共 R⬘ 兲 2 ⫽G̃ 共 R⬘ ←R, ␶ 兲 ⌿ T 共 R兲 2 ,
(3.51)
as can be verified from the definition G̃(R←R⬘ , ␶ )
⫽⌿ T (R)G(R←R⬘ , ␶ )⌿ T (R⬘ ) ⫺1 and the observation
that the exact G(R←R⬘ , ␶ ) is symmetric on interchange
of R and R⬘ . Second, if the trial function is equal to the
exact ground state, an attempt to impose this detailed
balance condition by accepting trial moves from R⬘ to R
with probability
p accept共 R←R⬘ 兲
冋
冋
G d 共 R⬘ ←R, ␶ 兲 G b 共 R⬘ ←R, ␶ 兲 ⌿ T 共 R兲 2
⫽min 1,
G d 共 R←R⬘ , ␶ 兲 G b 共 R←R⬘ , ␶ 兲 ⌿ T 共 R⬘ 兲 2
G d 共 R⬘ ←R, ␶ 兲 ⌿ T 共 R兲 2
⫽min 1,
G d 共 R←R⬘ , ␶ 兲 ⌿ T 共 R⬘ 兲 2
册
册
(3.52)
guarantees the correct sampling regardless of the size of
the time step, since the DMC algorithm reduces to the
VMC algorithm with trial moves sampled from G d . It
has been repeatedly demonstrated that the inclusion of
the acceptance/rejection step from Eq. (3.52) significantly decreases the time-step bias resulting from the
48
Foulkes et al.: Quantum Monte Carlo simulations of solids
approximate Green’s function. Obviously the acceptance ratio depends on the size of the time step and
tends to unity as the time step tends to zero, and the
approximate Green’s function more nearly satisfies the
detailed balance conditions of Eq. (3.51). If the time
step is large, however, and too many moves are rejected,
the introduction of the rejection step will bias the sampling of the unknown distribution f(R,t). The rule of
thumb is to choose the time step such that the acceptance ratio is ⭓99%. The better the approximate
Green’s function, the larger the time step that can be
used without a detectable bias (Umrigar, Nightingale,
and Runge, 1993).
Several other improvements to the Green’s function
are also normally employed. The use of an effective time
step in G b , as suggested by Reynolds et al. (1982) and
by Umrigar, Nightingale, and Runge (1993), further reduces the bias caused by preventing node crossing. As
mentioned above, the drift velocity diverges on the
nodal surface and so a walker close to a node can make
an excessively large move in the configuration space.
One can eliminate this problem, and thus improve the
approximate Green’s function, by imposing a cutoff on
the magnitude of the drift velocity. Similar cutoffs may
also be imposed on the values of the local energy used to
compute G b . Several different cutoff schemes have
been proposed (e.g., DePasquale, Rothstein, and Vrbik,
1988; Garmer and Anderson, 1988), but the smooth
forms suggested by Umrigar, Nightingale, and Runge
(1993) have been found to work well.
The result of the stochastic process described above is
a set of walker positions representing the distribution
f(R,t)⫽⌽(R,t)⌿ T (R). Given these positions, the expectation value of the energy can be evaluated in two
ways. One possibility is to measure the energy through
the offset E T that keeps the population constant on average. This is the so-called generational or populationgrowth estimator. In practice another possibility, known
as the mixed estimator, is normally used. This is given by
E D ⫽ lim
␶ →⬁
⫽ lim
␶ →⬁
⫽
具e
⌿ T 兩 Ĥ 兩 e
⫺ ␶ Ĥ/2
⌿ T兩 e
⫺ ␶ Ĥ/2
⫺ ␶ Ĥ/2
⌿ T典
⌿ T典
具 e ⫺ ␶ Ĥ ⌿ T 兩 Ĥ 兩 ⌿ T 典
具 e ⫺ ␶ Ĥ ⌿ T 兩 ⌿ T 典
具 ⌿ 0 兩 Ĥ 兩 ⌿ T 典
具 ⌿ 0兩 ⌿ T典
⫽ lim
␶ →⬁
⬇
具e
⫺ ␶ Ĥ/2
1
M
兰 f 共 R, ␶ 兲 E L 共 R兲 dR
兰 f 共 R, ␶ 兲 dR
兺m E L共 Rm 兲 ,
(3.53)
where 兵 Rm 其 is the set of M samples of f(R,⬁) resulting
from the DMC run. The accuracies and efficiencies of
these two energy estimators were compared by Umrigar,
Nightingale, and Runge (1993).
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Expectations of quantities that do not commute with
the Hamiltonian can be calculated using a combination
of the mixed and variational estimators
具 ⌽ 兩 Ŝ 兩 ⌽ 典 ⬇2 具 ⌽ 兩 Ŝ 兩 ⌿ T 典 ⫺ 具 ⌿ T 兩 Ŝ 兩 ⌿ T 典 ⫹O关共 ⌽⫺⌿ T 兲 2 兴 ,
(3.54)
where Ŝ is the operator corresponding to the physical
quantity of interest. Such combinations of VMC and
DMC estimators are often called extrapolated estimators. For nonnegative quantities (e.g., the density) one
can also use
具 ⌽ 兩 Ŝ 兩 ⌽ 典 ⬇ 具 ⌽ 兩 Ŝ 兩 ⌿ T 典 2 具 ⌿ T 兩 Ŝ 兩 ⌿ T 典 ⫺1 ⫹O关共 ⌽⫺⌿ T 兲 2 兴 .
(3.55)
The accuracy of an extrapolated estimator depends on
the trial function as well as the fixed-node DMC wave
function. This is a significant drawback, but alternative
approaches such as forward walking, which is clearly explained by Hammond, Lester, and Reynolds (1994), can
be used instead. Another possibility may be to use the
reptation QMC method of Baroni and Moroni (1999),
although this has not yet been tested in large systems. A
systematic discussion of the calculation of matrix elements of local, semilocal, and nonlocal operators that
either do or do not commute with the evolution operator is given by Nightingale (1999).
The most basic version of the DMC algorithm with
importance sampling (Reynolds et al., 1992; Hammond,
Lester, and Reynolds, 1994) can be outlined as follows:
(1) Generate a set of walkers drawn from some initial
distribution. (In most cases the initial walkers are
sampled from 兩 ⌿ T 兩 2 using variational Monte Carlo.)
Calculate the local energies of the walkers.
(2) Evaluate the drift velocity vD of each walker.
(3) Propagate each walker for a time step ␶, moving it
from its old position R⬘ to the new position
R⫽R⬘ ⫹ ␹ ⫹ ␶ vD 共 R⬘ 兲 ,
(3.56)
where ␹ is a 3N-dimensional vector of normally distributed numbers with variance ␶ and zero mean.
(4) Check whether the walker crossed the nodal surface (by checking the sign of the trial wave function) and
move it back to the original position if this is the case.
(5) Accept the step with the probability given by Eq.
(3.52).
(6) For each walker, calculate the number of copies
that will continue in the evolution using
M new⫽INT„␩ ⫹exp兵 ⫺ ␶ 关 E L 共 R兲 ⫹E L 共 R⬘ 兲 ⫺2E T 兴 /2其 …,
(3.57)
where ␩ is a random number drawn from a uniform distribution on the interval [0,1].
(7) Accumulate the quantities of interest. For example, the energy may be accumulated by averaging the
values of E L over the set of walkers.
(8) After an initial equilibration stage, steps 2–7 are
repeated until the error bars for averages of interest are
sufficiently small. The value of E T is occasionally adjusted to keep the average walker population roughly
constant. A simple formula for adjusting E T is
49
Foulkes et al.: Quantum Monte Carlo simulations of solids
E T ←E T ⫺C E ln共 M act /M ave兲 ,
(3.58)
where C E is a positive constant that controls how
quickly the actual number of walkers, M act , approaches
the desired number M ave . C E should not be too large
because this can bias the expectation values. However, it
cannot be too small because the number of walkers then
fluctuates too far from the desired value. Usually, it is
reasonable to choose a C E that rebalances the number
of walkers in 10 to 50 time steps. More sophisticated
techniques for dealing with E T are discussed by Umrigar, Nightingale, and Runge (1993).
IV. TRIAL WAVE FUNCTIONS
A. Introduction
The quality of the trial wave function controls the statistical efficiency and limits the final accuracy of any
VMC or DMC simulation. The repeated evaluation of
the trial wave function (and its gradient and Laplacian)
is also the most demanding part of the computation. We
therefore seek trial wave functions that are both accurate and easy to evaluate.
In quantum-chemical methods it is common to express many-body wave functions as linear combinations
of determinants. However, such expansions converge
very slowly because of the difficulty in describing the
cusps that occur whenever two electrons come into contact. QMC simulations of solids require a much more
compact representation and normally use trial wave
functions of the Slater-Jastrow type (Jastrow, 1955),
consisting of a single Slater determinant multiplied by a
totally symmetric non-negative Jastrow correlation factor that includes the cusps. The orbitals in the Slater
determinant are usually obtained from accurate LDA or
HF calculations, while the Jastrow factor is chosen to
have some specific functional form and optimized as explained in Sec. VII. Linear combinations of SlaterJastrow functions are sometimes required, but the
single-determinant form is satisfactory for many purposes. When used in VMC simulations of tetrahedral
semiconductors (Fahy, Wang, and Louie, 1988, 1990a;
Li, Ceperley, and Martin, 1991; Rajagopal, Needs,
Kenny et al., 1994; Rajagopal et al., 1995; Eckstein et al.,
1996; Malatesta, Fahy, and Bachelet, 1997), singledeterminant Slater-Jastrow trial functions generally reproduce the experimental cohesive energies to within
about 0.1 eV per atom.
B. Slater-Jastrow wave functions
A single-determinant Slater-Jastrow wave function
can be written as
⌿ 共 X兲 ⫽e J(X) D 共 X兲 ,
(4.1)
where D(X) is a Slater determinant as shown in Eq.
(2.3), X⫽(x1 ,x2 ,...,xN ), and xi ⫽ 兵 ri , ␴ i 其 denotes the
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
FIG. 6. Spin-parallel and spin-antiparallel pair-correlation
functions calculated within the Hartree-Fock (HF) approximation for a uniform three-dimensional electron gas. The HF
pair-correlation functions depend only on the product of the
Fermi wave vector k f and the interparticle separation r.
space and spin coordinates of electron i. Most simulations retain only one- and two-body terms in the Jastrow
factor,
N
J 共 X兲 ⫽
兺
i⫽1
N
␹ 共 xi 兲 ⫺
1
2 i⫽1
N
兺 兺
u 共 xi ,xj 兲 .
(4.2)
j⫽1
(j⫽i)
The u terms describe the electron-electron correlations,
while the ␹ function depends on the positions of the
nuclei and describes the electron-nuclear correlation.
In QMC calculations one normally removes the spin
variables from the Slater-Jastrow wave function and replaces the single Slater determinant of Eq. (4.1) by a
product of determinants of up-spin and down-spin orbitals,
⌿ 共 R兲 ⫽e J(R) D ↑ 共 r1 ,...,rN ↑ 兲 D ↓ 共 rN ↑ ⫹1 ,...,rN 兲 ,
(4.3)
where R⫽(r1 ,r2 ,...,rN ) denotes the spatial coordinates
of all the electrons. This function is not antisymmetric
on exchange of electrons with opposite spins and so differs from the Slater-Jastrow wave function of Eq. (4.1),
but in Sec. IV.E we show that it gives the same expectation value for any spin-independent operator. The use
of spin-independent wave functions such as ⌿(R) is
computationally advantageous because a large determinant is replaced by two smaller ones and no sums over
spin variables are required; it also facilitates the imposition of the cusp conditions discussed in Sec. IV.F. By
convention, we shall always choose r1 ,...,rN ↑ to be the
coordinates of the spin-up electrons and rN ↑ ⫹1 ,...,rN to
be the coordinates of the spin-down electrons.
The physics underlying Slater-Jastrow wave functions
is quite straightforward. Figure 6 shows the spin-parallel
and spin-antiparallel pair-correlation functions of a
three-dimensional uniform electron gas calculated
within the HF approximation. The antisymmetry of the
wave function creates an exchange hole that keeps
parallel-spin electrons apart, but there are no correla-
50
Foulkes et al.: Quantum Monte Carlo simulations of solids
FIG. 7. Charge densities of a full core carbon atom from an
LDA calculation (solid line) and a Slater-Jastrow wave function with a two-body correlation term but no one-body correlation term (dotted line). We thank Mark Stedman for providing this figure.
tions between the positions of electrons with antiparallel
spins. There is a significant probability of finding two
antiparallel-spin electrons very close to each other and
so the electron-electron Coulomb repulsion energy is
high. The purpose of the two-body u term in the Jastrow
factor is to reduce the magnitude of the many-electron
wave function whenever two electrons approach one another; this reduces the probability of finding two electrons very close together and decreases the electronelectron interaction energy.
In uniform systems this is all that is required. In nonuniform systems, however, the introduction of the repulsive u term changes the electron density significantly,
pushing electrons away from regions of high charge density and into low-density regions. This is an unavoidable
consequence of introducing a repulsive two-body term,
even though the intention was to change the paircorrelation function, not the density. In fact, the charge
densities of Slater determinants of HF or LDA orbitals
are usually quite accurate, and the change caused by the
u term is unwelcome. The one-body term in Eq. (4.2)
alters the charge density without greatly disturbing the
pair-correlation function and hence allows this problem
to be overcome.
The effect of the ␹ term is illustrated in Figs. 7 and 8,
which show the spherically averaged charge densities of
a full core carbon atom. Figure 7 shows that the LDA
charge density (solid line) is significantly altered by the
introduction of a two-body u function (dotted line). The
u function lowers the energy only slightly because the
change in the charge density is unfavorable. Figure 8
shows that the charge density (dotted line) returns almost to the LDA form (solid line again) if a one-body ␹
function is added to the wave function and optimized
variationally. It is the combination of optimized ␹ and u
terms which lowers the calculated atomic energy greatly.
Similar behavior is observed in simulations of solids, although with some quantitative differences when the atRev. Mod. Phys., Vol. 73, No. 1, January 2001
FIG. 8. Charge densities of a full core carbon atom from an
LDA calculation (solid line) and a Slater-Jastrow wave function with both one- and two-body correlation terms (dotted
line). We thank Mark Stedman for providing this figure.
oms are represented using pseudopotentials. The introduction of the u term alone typically lowers the energy
per atom by an eV but suppresses the charge density in
the bonding regions and enhances it in the interstitial
regions. The introduction of the ␹ term restores the
charge density to something close to the LDA form and
lowers the energy by an additional fraction of an eV.
The importance of the one-body ␹ term was recognized long ago in the context of Fermi-hypernettedchain calculations (Feenberg, 1969) and in the correlation factor introduced by Boys and Handy (1969). The
introduction of ␹ functions into QMC simulations of solids was due to Fahy et al. (1988, 1990a). Although the e ␹
factors can be absorbed into the single-particle orbitals
of the Slater determinant, they are so closely linked to
the u function that it is normally convenient to keep
them in the Jastrow factor. The relationship between the
u and ␹ functions is discussed in more detail in Sec.
IV.D.
The great success of single-determinant Slater-Jastrow
wave functions is founded on the success of the HF approximation. For example, in the carbon atom HF
theory retrieves 99.6% of the total energy. The other
0.4% (4.3 eV) is the correlation energy, which is important for an accurate description of chemical bonding but
is small compared with the total energy. Quantum
Monte Carlo calculations for solids almost always use
pseudopotentials in which the large energy contribution
from the core electrons is removed. In this case the correlation energy can amount to several percent of the
total energy, but the HF approximation still accounts for
the vast majority of the energy. For example, in the carbon pseudo-atom HF theory retrieves about 98.2% of
the total energy. The error of 1.8% (2.7 eV) is a small
fraction of the total but is large on the scale of bonding
energies. Even a relatively simple Jastrow factor can
lower the total energy of a carbon pseudo-atom by
about 2.4 eV, thereby retrieving 89% of the correlation
energy and reducing the error in the total energy to
Foulkes et al.: Quantum Monte Carlo simulations of solids
0.2%. The Jastrow factor therefore provides a small but
very important correction to the wave function.
51
coordinates ri appearing in the Slater determinants of
Eq. (4.3) by the corresponding quasiparticle coordinates
N
r̄i ⫽ri ⫹
C. The Slater determinant
The orbitals for the Slater determinant are most commonly obtained from HF or LDA calculations. This procedure may be motivated by noting that a determinant
of HF orbitals gives the lowest energy of all singledeterminant wave functions (although it is far from obvious that HF or LDA orbitals remain a good choice in
the presence of a Jastrow factor). Here we consider the
few attempts that have been made to devise something
better.
Direct numerical optimization of the single-particle
orbitals in the Slater determinant has a long history in
quantum chemistry but has not often been attempted
using QMC methods. To date, most of the work on stochastic orbital optimization has been for fairly small systems (see, for example, Umrigar, Nightingale, and
Runge, 1993), although Eckstein and co-workers successfully optimized the parameters of the very restricted
Slater orbital basis sets used in their studies of solid Li
(Eckstein and Schattke, 1995) and GaAs (Eckstein et al.,
1996). More recently, Fahy (1999) and Filippi and Fahy
(2000) have developed a new method that makes it possible to optimize the more accurate single-particle states
used in most current QMC simulations of solids. Although LDA or HF orbitals are usually adequate for
real solids, direct numerical optimization will probably
become widespread within a few years.
A different tack was taken by Grossman and Mitas
(1995), who used a determinant of the natural orbitals
which diagonalize the one-electron density matrix. The
motivation is that the convergence of configurationinteraction expansions is improved by using natural orbitals instead of HF orbitals. Grossman and Mitas (1995)
obtained the natural orbitals for Si2, Si3, and Si4 molecules via multiconfiguration HF calculations. They
found that the use of natural orbitals lowered the QMC
energy significantly, but a similar calculation for bulk
silicon (Kent et al., 1998) found little improvement. Further work revealed that for wave functions with little
mixing of excited configurations the decrease in energy
was rather small. However, for more complicated wave
functions the improvements were significant, and natural
orbitals proved necessary to obtain systematically high
accuracy in calculations of molecular reactions (Grossman and Mitas, 1997).
Another approach is based on the idea of backflow
correlations, derived from a current-conservation argument by Feynman and Cohen (1956) to provide a picture of the excitations in liquid 4He. Work on liquid 3He
(Schmidt et al., 1981; Panoff and Carlson, 1989), various
atoms (Schmidt and Moskowitz, 1992), and the uniform
electron gas (Kwon, Ceperley, and Martin, 1993, 1994,
1998) shows that backflow correlations are also helpful
in fermionic systems. The essential ingredient of the
backflow trial function is the replacement of the electron
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
兺
j⫽1
(j⫽i)
␩ 共 r ij 兲共 ri ⫺rj 兲 ,
(4.4)
where r ij ⫽ 兩 ri ⫺rj 兩 . In uniform systems, this amounts to
replacing the single-particle orbitals exp(ikl •ri ) by
exp(ikl •r̄i ), which ensures that the correlation between
electrons i and j depends on the momentum kl of the
state occupied by electron i. The optimal function ␩ (r ij )
is usually determined variationally. Kwon, Ceperley, and
Martin (1998) found that the introduction of backflow
significantly lowered the VMC and DMC energies of the
three-dimensional uniform electron gas at high densities
(the VMC energy decreased by 0.07 eV per electron
when r s ⫽1 and by 0.02 eV per electron when r s ⫽5).
The same authors (Kwon, Ceperley, and Martin, 1993,
1998) also investigated the importance of three-electron
correlation terms in the Jastrow factor, obtaining a small
improvement in the energy at low densities. The use of
backflow wave functions for inhomogeneous systems is
much less explored, but since backflow increases the
computational cost of QMC calculations significantly its
use will probably be confined to cases in which extremely high accuracy is sought.
In some cases it is necessary to use multideterminant
wave functions to preserve important symmetries of the
true wave functions. In other cases a single determinant
may give the correct symmetry, but a significantly more
accurate wave function can be obtained by using a linear
combination of a few determinants. For example, in the
Be atom the 2s and 2p orbitals are nearly degenerate
and the many-electron wave function contains important
contributions from the electronic configurations 1s 2 2s 2
and 1s 2 2p 2 (Harrison and Handy, 1985). The accurate
variational trial function for Be proposed by Umrigar,
Nightingale, and Runge (1993) is a linear combination of
the corresponding Slater-Jastrow wave functions. Flad,
Caffarel, and Savin (1997) have carried out a thorough
investigation of the advantages and disadvantages of using multideterminant wave functions for various other
atoms. In solids, however, the number of important configurations may grow very rapidly with system size and
the use of multireference wave functions can be very
demanding.
D. The Jastrow factor
What is an appropriate Jastrow factor for a uniform
electron gas? The spatial homogeneity of the system ensures that no ␹ terms are needed and that u depends
only on r ij ⫽ 兩 ri ⫺rj 兩 . The cusp conditions (Kato, 1957;
Pack and Brown, 1966; and Sec. IV.F) determine the
behavior of the u function in the limit r ij →0, while arguments based on the random-phase approximation
(RPA) of Bohm and Pines (1953) imply that it has a 1/r ij
tail when r ij is large.
The following simple choice has the correct small- and
large-r ij limits and was used (along with a better form
52
Foulkes et al.: Quantum Monte Carlo simulations of solids
based more closely on the RPA) in some of the earliest
work on the homogeneous electron gas (Ceperley, 1978;
Ceperley and Alder, 1980):
u ␴ i , ␴ j 共 ri ,rj 兲 ⫽
A ␴i ,␴j
r ij
共 1⫺e ⫺r ij /F ␴ i , ␴ j 兲 .
(4.5)
The correlations between pairs of electrons depend on
whether they have parallel or antiparallel spins, so the
constants A ␴ i , ␴ j and F ␴ i , ␴ j are spin dependent. Assuming that the solid is not spin polarized, there are four
parameters: A ↑↑ ⫽A ↓↓ , A ↑↓ ⫽A ↓↑ , F ↑↑ ⫽F ↓↓ , and F ↑↓
⫽F ↓↑ . Two of the four may be eliminated by imposing
the cusp conditions
du ␴ i , ␴ j 共 r ij 兲
dr ij
冏 再
⫽
r ij ⫽0
⫺1/4 ␴ i ⫽ ␴ j ,
⫺1/2 ␴ i ⫽ ␴ j .
(4.6)
Section IV.F shows that as long as the cusp conditions
are obeyed the divergence in the Coulomb interaction
energy occurring whenever two electrons coincide is exactly cancelled by a divergence in the kinetic energy
arising from the cusp in the wave function. For the Jastrow factor of Eq. (4.5) the cusp conditions translate
into
A ↑↑
2
2F ↑↑
⫽
1
4
and
A ↑↓
2
2F ↑↓
1
⫽ .
2
(4.7)
The remaining two parameters, which we take to be A ↑↑
and A ↑↓ , can be chosen to reproduce the long-range
behavior implied by the RPA,
u ␴ i , ␴ j 共 r ij 兲 ⬇
1
as r ij →⬁,
␻ p r ij
(4.8)
where ␻ p ⫽ 冑4 ␲ n is the plasma frequency and n is the
electron number density. This implies that
A ↑↑ ⫽A ↑↓ ⫽
1
1
⫽
.
␻ p 冑4 ␲ n
(4.9)
An alternative is to optimize the values of A ↑↑ and A ↑↓
variationally, but this does not give much improvement
in practice. Most solid-state QMC calculations these
days use Eq. (4.5) or something similar as a starting
point and add variational degrees of freedom to improve
the accuracy. The optimal values of the variational parameters are determined using the techniques discussed
in Sec. VII.
The cusp conditions imply that the Jastrow factor is
spin dependent and this produces a small amount of spin
contamination: a linear combination of Slater-Jastrow
trial functions may not be an eigenfunction of Ŝ 2 even
though the corresponding sum of Slater determinants is.
It is difficult to clean up this contamination without
spoiling the accuracy of the trial function in other ways,
but its effects are so small that they can almost always be
ignored (Huang, Filippi, and Umrigar, 1998).
In QMC calculations we model extended systems using small simulation cells subject to periodic boundary
conditions. This periodicity must be reflected in
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
u ␴ ⬘ , ␴ (r⬘ ,r), which should remain unchanged if either r⬘
or r is translated by a simulation-cell lattice vector. One
way to enforce the correct translational symmetry on the
u function of Eq. (4.5) is to sum it over all the periodic
‘‘images’’ of the electrons in an infinite lattice of identical copies of the simulation cell. The resulting sum does
not converge because u has a slowly decaying 1/r ij tail,
but can be made meaningful by adopting the Ewald definition (Ewald, 1921; see Tosi, 1964, for a clear review).
Ortiz and Ballone (1994) and Williamson et al. (1996)
have used u functions that tend to zero smoothly as r ij
approaches the radius r WS of the largest sphere that fits
inside the Wigner-Seitz simulation cell. This is computationally efficient because the need for sums over images
is eliminated. However, since u is set to zero for r ij
⬎r WS , it no longer has the 1/r ij tail implied by the RPA.
In practice, the consequences of truncating u appear to
be small for reasonably large simulation cells.
A suitable form for the ␹ function in a periodic solid is
a plane-wave expansion,
␹ ␴ 共 r兲 ⫽
␹ ␴ ,G e iG •r,
兺
G
p
p
p
(4.10)
where Gp is a reciprocal vector of the primitive crystal
lattice (not the lattice of simulation cells, each of which
may contain several primitive unit cells). This form has
been used by Fahy and co-workers in studies of sp
bonded materials (Fahy, Wang, and Louie, 1988, 1990a).
In Fahy’s original work (Fahy, Wang, and Louie, 1988,
1990a) the ␹ function was designed to force the electron
density towards the LDA result, which was assumed to
be accurate. This ad hoc procedure is fairly successful,
but has been superseded by direct optimization of the
␹ ␴ ,Gp coefficients (Williamson et al., 1996; Malatesta,
Fahy, and Bachelet, 1997). The number of independent
variational parameters may be reduced by noting that
Gp vectors of the same length are often related by symmetry. When full use is made of such symmetry arguments, only six independent parameters are required to
produce an accurate ␹ function for crystalline Ge (Williamson et al., 1996).
In calculations for finite systems or for solids with
many inequivalent atoms the plane-wave representation
of ␹ is inefficient. To overcome this problem one can
represent ␹ by a sum of local atom-centered functions,
␹ ␴ 共 r兲 ⫽
兺␣ g ␴ , ␣共 兩 r⫺d␣兩 兲 .
(4.11)
This form proved useful in the study of the formation
energies of silicon self-interstitial defects discussed in
Sec. V.I. Since the atoms around the defect were not all
equivalent, it was found beneficial to allow the ␹ functions on inequivalent atoms to differ.
An important advance in understanding Jastrow factors was made by Malatesta, Fahy, and Bachelet (1997;
see also Fahy, 1999). The discussion in Sec. IV.B suggests that the one- and two-body terms in the Jastrow
factor are closely related, and we might expect that a
linear relationship exists between the optimal forms of u
53
Foulkes et al.: Quantum Monte Carlo simulations of solids
and ␹. Assuming that the u function depends only on
the relative positions of the two electrons, u ␴ i , ␴ j (ri ,rj )
⫽u ␴ i , ␴ j (ri ⫺rj ), the relationship derived by Malatesta,
Fahy, and Bachelet (1997) takes the form
␹ ␴ ,Gp ⫽
具 n̂ ␴*⬘ ,G 典 u ␴ ⬘ , ␴ ,G
兺
␴
p
p
⬘
,
(4.12)
where ␹ ␴ ,Gp is a Fourier component of ␹ ␴ (r) as defined
in Eq. (4.10), u ␴ ⬘ , ␴ ,Gp is the corresponding Fourier component of u ␴ ⬘ , ␴ (r), and the operators
N↑
n̂ ↑,Gp ⫽
兺e
i⫽1
⫺iGp •ri
N ↑ ⫹N ↓
and n̂ ↓,Gp ⫽
兺
i⫽N ↑ ⫹1
e ⫺iGp •ri
(4.13)
measure the Fourier components of the up- and downspin electron number densities [multiplied by the
simulation-cell volume if the Fourier components are
defined in analogy with Eq. (4.10)]. Note that unlike
␹ ␴ (r), which has the full periodicity of the crystal unit
cell, the u function has only the periodicity of the simulation cell. The Fourier series for u ␴ ⬘ , ␴ (r) therefore contains Fourier components for every vector Gs in the
simulation-cell reciprocal lattice. The operator n̂ ␴ ,Gs is
also defined for all simulation-cell reciprocal-lattice vectors Gs , but its expectation value 具 n̂ ␴ ,Gs 典 vanishes unless
Gs 苸 兵 Gp 其 . The set 兵 Gp 其 is a subset of 兵 Gs 其 .
Malatesta, Fahy, and Bachelet (1997) used Eq. (4.12)
with the LDA charge density and the u function of Eq.
(4.5) to construct a ␹ function for a VMC calculation of
cubic boron nitride. The result was almost, but not quite,
as good as a ␹ function obtained (at much greater cost)
using the variance minimization method explained in
Sec. VII. There is therefore little doubt that Eq. (4.12)
provides a valuable insight into the physics of Jastrow
functions as well as a practical way of obtaining useful
approximations to the ␹ function.
The reasoning behind Eq. (4.12) was based on the
Bohm-Pines (1953) RPA theory of the uniform electron
gas, which leads to a spin-independent Jastrow factor. If,
for the sake of convenience, we choose to add constant
diagonal (i⫽j) terms to the Jastrow factor, we can express it in terms of Fourier components as follows:
N
1
⫺
2 i⫽1
N
兺 j⫽1
兺
兺
G
s
* 典uG .
␹ Gp ⫽ 具 n̂ G
p
p
(4.16)
Equation (4.12) is the natural generalization of Eq.
(4.16) to the case when the Jastrow factor is spin dependent.
In real materials, u ␴ i , ␴ j (ri ,rj ) no longer depends only
on r ij but on ri and rj separately. In atomic and molecular calculations it has become common to take account
of this inhomogeneity by including three-body electronelectron-nucleus correlation terms. The Jastrow factor is
then of the form
J 共 R兲 ⫽
u ␴ , ␴ 共 r i ␣ ,r j ␣ ,r ij 兲 ,
兺␣ 兺
i,j
i
(4.17)
j
where r i ␣ is the separation of the ith electron from the
␣ th nucleus. Schmidt and Moskowitz (1992) have used
the correlation factor of Boys and Handy (1969), while
Umrigar and co-workers (Umrigar, Wilson, and Wilkins,
1988; Filippi and Umrigar, 1996; Huang, Filippi, and
Umrigar, 1998) have used various Padé and polynomial
forms. According to Huang, Umrigar, and Nightingale
(1997), the inclusion of electron-electron-nucleus correlation terms lowers the VMC energy of the Be atom by
0.16 eV. The same authors also considered four-body
electron-electron-electron-nucleus terms, which improved the VMC energy but not the DMC energy, even
though the determinantal part of the trial function was
reoptimized along with the Jastrow factor. Schmidt and
Moskowitz (1992) have shown that the electronelectron-nucleus terms can describe some of the effects
of backflow corrections. The inhomogeneity of the Jastrow factor may be less important in pseudopotential
calculations, but tests for 3d transition-metal atoms (Mitas, 1993) indicate that it can still be significant.
E. Spin
1
u 共 ri ⫺rj 兲 ⫽⫺
2
* u G n̂ G .
n̂ G
兺
G
s
s
s
(4.14)
s
The insight of Malatesta, Fahy, and Bachelet (1997) was
to see that in extending the RPA to inhomogeneous systems, one should replace the Fourier components of the
electron number densities by the deviations from their
mean values, ⌬n̂ Gs ⫽n̂ Gs ⫺ 具 n̂ Gs 典 . The Jastrow factor
then becomes (Malatesta, Fahy, and Bachelet, 1997)
1
⫺
2
where we have assumed that u(r) is real and used our
earlier observation that 具 n̂ Gs 典 is zero unless Gs 苸 兵 Gp 其 .
The right-hand side contains a two-body term of the
form of Eq. (4.14) and a one-body term with Fourier
components,
1
* u G ⌬n̂ G ⫽⫺
⌬n̂ G
s
s
s
2
⫹
* u G n̂ G
n̂ G
兺
G
s
s
s
具 Ô 典 ⫽
s
* 典 u G n̂ G ⫹const,
具 n̂ G
兺
G
p
p
p
p
(4.15)
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
At first sight it may appear that the spin dependence
of trial functions such as ⌿(X) from Eq. (4.1) complicates QMC algorithms considerably. This is not the case,
as we now explain. Suppose we wish to use ⌿(X) to
calculate the expectation value of a spin-independent
operator Ô(R), where R⫽(r1 ,r2 ,...,rN ):
兺␴
冕
⌿ * 共 X兲 Ô 共 R兲 ⌿ 共 X兲 dR
兺␴
⌿ * 共 X兲 ⌿ 共 X兲 dR
冕
.
(4.18)
For each spin configuration ␴⫽( ␴ 1 , ␴ 2 ,..., ␴ N ), we can
replace the totally antisymmetric wave function
⌿(x1 ,x2 ,...,xN ) by a version with permuted arguments
54
Foulkes et al.: Quantum Monte Carlo simulations of solids
⌿(xi 1 ,xi 2 ,...,xi N ), choosing the permutation such that
the first N ↑ arguments are spin up and the remaining
N ↓ ⫽N⫺N ↑ are spin down:
⌿ 共 x1 ,...,xN 兲 →⌿ 共 xi 1 ,...,xi N 兲
⫽⌿ 共 兵 ri 1 ,↑ 其 ,..., 兵 ri N ,↑ 其 , 兵 ri N
↑
↑ ⫹1
,↓ 其 ,..., 兵 ri N ,↓ 其 兲 .
(4.19)
Since R is a dummy variable of integration, we can now
relabel (ri 1 ,ri 2 ,...,ri N ) as (r1 ,r2 ,...,rN ),
⌿ 共 兵 ri 1 ,↑ 其 ,..., 兵 ri N ,↑ 其 , 兵 ri N
↑
↑ ⫹1
,↓ 其 ,..., 兵 ri N ,↓ 其 兲
→⌿ 共 兵 r1 ,↑ 其 ,..., 兵 rN ↑ ,↑ 其 , 兵 rN ↑ ⫹1 ,↓ 其 ,..., 兵 rN ,↓ 其 兲 .
(4.20)
The operator Ô is not affected by the relabelling because it is symmetric with respect to exchange of electrons and so each spin configuration gives an identical
contribution to the expectation value. We can therefore
remove the sums over spin configurations to get
具 Ô 典 ⫽
兰 ⌿ * 共 R兲 Ô 共 R兲 ⌿ 共 R兲 dR
,
兰 ⌿ * 共 R兲 ⌿ 共 R兲 dR
(4.21)
⌿ 共 R兲 ⫽⌿ 共 r1 ,...,rN 兲
⫽⌿ 共 兵 r1 ,↑ 其 ,..., 兵 rN ↑ ,↑ 其 , 兵 rN ↑⫹1 ,↓ 其 ,..., 兵 rN ,↓ 其 兲 .
(4.22)
The new wave function ⌿(R) gives the same expectation values as ⌿(X) but is much more convenient to use
in QMC calculations.
This analysis allows us to replace the spin-dependent
trial state ⌿(X) by a spin-independent state ⌿(R). The
new state is antisymmetric with respect to exchange of
the spatial coordinates of pairs of spin-up electrons or
pairs of spin-down electrons, but has no specific symmetry on exchange of the spatial coordinates of electrons
with different spins. We are therefore treating spin-up
electrons as distinguishable from spin-down electrons.
We can reconstruct ⌿(X) by multiplying ⌿(R) by the
spin function
↑
↑ ⫹1
,↓ ... ␦ ␴ N ,↓
(4.23)
and antisymmetrizing.
F. The cusp conditions
We have made repeated reference to the cusp conditions (Kato, 1957; Pack and Brown, 1966), which we now
discuss in more detail. Consider the Schrödinger equation for the ground state of the hydrogen atom,
1
1
Ĥ ␺ 1s ⫽⫺ ⵜ 2 ␺ 1s ⫺ ␺ 1s ⫽E 1s ␺ 1s .
2
r
(4.24)
The potential energy diverges as r→0 but Ĥ ␺ 1s remains
finite because there is a cancelling divergence in the kiRev. Mod. Phys., Vol. 73, No. 1, January 2001
Ĥ⌿ 0 共 r1 ,r2 ,...,rN 兲 ⫽E 0 ⌿ 0 共 r1 ,r2 ,...,rN 兲 ,
(4.25)
where
N
Ĥ⫽⫺
where the spatial wave function ⌿(R) is defined by
␦ ␴ 1 ,↑ ... ␦ ␴ N ,↑ ␦ ␴ N
netic energy. This is reflected in the shape of the 1s
eigenfunction, which has a sharp cusp at the origin.
The potential energy diverges whenever an electron
approaches a nucleus (as long as that nucleus is not represented by a pseudopotential) or another electron.
Many-electron wave functions therefore contain similar
cusps. Any necessary electron-nuclear cusps are built
into the single-particle orbitals obtained by solving the
mean-field HF or LDA equations, but these orbitals do
not contain the electron-electron cusps. The purpose of
applying the cusp conditions (Kato, 1957; Pack and
Brown, 1966) stated in Eq. (4.6) is to build these missing
cusps into the Jastrow factor. Experience shows that enforcing the cusp conditions reduces both the average energy and its variance.
The many-electron ground state ⌿ 0 (R) satisfies the
Schrödinger equation
1
ⵜ 2 ⫹V 共 r1 ,r2 ,...,rN 兲 .
2 i⫽1 ri
兺
(4.26)
If we pick a pair of electrons i and j and introduce the
difference and center-of-mass variables r⫽ri ⫺rj and
rc.m.⫽(ri ⫹rj )/2, Ĥ may be rewritten as
1
1
Ĥ⫽⫺ⵜ r2 ⫺ ⵜ r2 ⫺
4 c.m. 2
N
兺
k⫽1
(k⫽i,j)
ⵜ r2 ⫹V 共 r1 ,r2 ,r3 ,...,rN 兲 .
k
(4.27)
Since ⌿ 0 is an exact eigenfunction, the corresponding
exact local energy E L0 ⫽⌿ ⫺1
0 Ĥ⌿ 0 is everywhere equal
to the ground-state energy E 0 and does not diverge as
r→0.
The Slater-Jastrow trial function ⌿ is not an exact
eigenfunction and so the local energy E L ⫽⌿ ⫺1 Ĥ⌿ is
not constant and may diverge as r→0. To study this possibility we write ⌿ in the form
⌿⫽e ⫺u(r) f 共 r兲 ,
(4.28)
where f(r) includes the Slater determinants and everything from the Jastrow factor except the term involving
u ␴ i , ␴ j (r). Although f(r) depends on all the variables
needed to describe the system, only the dependence on r
has been shown explicitly; the spin labels of u ␴ i , ␴ j (r)
have also been suppressed. For the sake of simplicity, we
shall assume that u is a function of r only and does not
depend on ri and rj separately, although the cusp conditions apply more generally.
Consider the behavior of the local energy
1
1
Ĥe ⫺u(r) f 共 r兲
Ĥ⌿⫽ ⫺u(r)
⌿
e
f 共 r兲
(4.29)
as r tends to zero while rc.m. and all the other electron
positions are held fixed. For almost every possible
choice of the fixed coordinates, the function f(r) and all
its derivatives are finite as r→0, so any divergent terms
Foulkes et al.: Quantum Monte Carlo simulations of solids
in the kinetic energy must arise from the action of the
ⵜ r2 operator on e ⫺u(r) . The local energy therefore remains finite as long as
1
e
⫺u
f
冉
⫺ⵜ 2 ⫹
⫽u ⬙ ⫹
冊
1
共 e ⫺u f 兲
r
2u ⬘
ⵜf ⵜ 2 f 1
⫺ 共 u ⬘ 兲 2 ⫹2u ⬘ r̂•
⫺
⫹
r
f
f
r
(4.30)
remains finite, where r̂ is a unit vector in the r direction
and the primes denote differentiation with respect to r.
If electrons i and j have opposite spins, the value of
f(r⫽0) is in general nonzero. We insist that u ⬘ and u ⬙
tend to finite values as r→0, and so only the second and
final terms in Eq. (4.30) diverge. These two divergences
cancel if we impose the opposite-spin cusp condition
du
dr
冏
1
⫽⫺ .
2
r⫽0
(4.31)
If electrons i and j have parallel spins, the Pauli principle ensures that f(r) is an odd function of r. It therefore has a Taylor expansion of the form
f 共 r兲 ⫽a•r⫹O共 r 3 兲 ,
(4.32)
where a⫽ⵜf(r) 兩 r⫽0 . It follows from the form of this series (which is assumed to have a finite radius of convergence) that the Laplacian of f tends to zero as r→0, but
the second, fourth, and last terms in Eq. (4.30) all diverge. The sum of the divergent contributions is
2u ⬘ 2u ⬘
1 4u ⬘ ⫹1
⫹
r̂•a⫹ ⫽
,
r
a•r
r
r
(4.33)
and so the parallel-spin cusp condition is
du
dr
冏
1
⫽⫺ .
4
r⫽0
(4.34)
V. SELECTED APPLICATIONS OF QUANTUM
MONTE CARLO TO GROUND STATES
A. Cohesive energies of solids
The cohesive energy of a solid is the difference between the energy of a dispersed gas of the constituent
atoms or molecules and the energy of the solid at zero
temperature. Cohesive energies can be obtained by
measuring the latent heat of sublimation and extrapolating to zero temperature. Calculating a cohesive energy is
a severe test of QMC techniques because it is the energy
difference between two very different systems. To obtain an accurate cohesive energy within VMC one must
use trial wave functions of closely matching accuracy for
the atom and solid. DMC calculations are less sensitive
to the trial wave function but the quality of the fixed
nodal surfaces must still be comparable.
In the pioneering work of Fahy, Wang, and Louie
(1988, 1990a), VMC methods with nonlocal pseudopotentials were used to calculate cohesive energies of
diamond-structure C and Si. The calculations were perRev. Mod. Phys., Vol. 73, No. 1, January 2001
55
formed with simulation cells containing 16 and 54 atoms
and used trial wave functions of the Slater-Jastrow type.
The VMC results were surprisingly accurate, giving cohesive energies within about 0.2 eV of the experimental
values. This success was followed by a VMC and DMC
study of diamond-structure Si by Li, Ceperley, and Martin (1991). This calculation represented another significant step forward because it was the first application of
the DMC method to a heavy-atom solid. The DMC cohesive energy of Si obtained by Li, Ceperley, and Martin
(1991) differs from experiment by only about 0.1 eV,
although a correction of about 0.2 eV was applied to
compensate for an error introduced by using a pseudoHamiltonian of the type discussed in Sec. VIII.F to
eliminate the core electrons. Another important advance was the development of the locality approximation for evaluating the energy of a nonlocal pseudopotential within DMC calculations (Mitas, Shirley, and
Ceperley, 1991). This scheme has significantly increased
the scope of DMC calculations and has made it possible
to perform accurate DMC calculations of the cohesive
energies of heavy-atom solids.
QMC calculations of the cohesive energies of a number of solids have been performed, and in Table I we
have compiled the available results on the tetrahedrally
bonded semiconductors, C, Si, Ge, and cubic BN. The
calculated results include corrections for the finite sizes
of the simulation cells and for the zero-point motion of
the nuclei. We refer the reader to the original references
for details of these corrections. The finite-size corrections are particularly important and were very carefully
studied by Kent, Hood, et al. (1999). Diffusion Monte
Carlo results for cells containing up to 250 atoms were
used to obtain a cohesive energy for Si of 4.63(2) eV per
atom, which is very close to the experimental value of
4.62(8) eV per atom. This indicates that the DMC
method can give cohesive energies to better than 0.1 eV
per atom in sp-bonded systems. The data in Table I
show the well-known trend that the cohesive energies
calculated within the local spin-density approximation
(LSDA) are too large in sp-bonded systems. The largest
contribution to the error arises from the atom; for example, in silicon the LSDA energy is about 1.1 eV per
atom too high in the pseudo-atom and about 0.4 eV too
high in the pseudosolid, resulting in an overestimation of
the cohesive energy by about 0.7 eV.
Very few groups have as yet applied QMC methods to
more strongly correlated real solids, but examples include the study of NiO by Tanaka (1993) and a study of
the Fe atom by Mitas (1994). The results so far suggest
that accuracies better than 1 eV per atom can be obtained. This is still a far cry from the accuracy required
to study high-temperature superconductivity, for example, which would need an improvement of at least
two orders of magnitude. The major difficulty is to construct trial wave functions that capture the essence of
the physics of such complicated systems.
B. Phases of the electron gas
The homogeneous electron gas is the simplest realistic
model of interacting electrons, and studies of its proper-
56
Foulkes et al.: Quantum Monte Carlo simulations of solids
TABLE I. Cohesive energies of tetrahedrally bonded semiconductors calculated within the local
spin-density approximation (LSDA), variational Monte Carlo (VMC), and diffusion Monte Carlo
(DMC) methods and compared with experimental values. The energies for Si, Ge, and C are quoted
in eV per atom, while those for BN are in eV per two atoms.
Method
Si
Ge
C
BN
LSDA
VMC
5.28a
4.38(4)c
4.82(7)d
4.48(1)g
4.51(3)c
4.63(2)g
4.62(8)a
4.59a
3.80(2)b
8.61a
7.27(7)d
7.36(1)f
15.07e
12.85(9)e
3.85(2)b
7.46(1)f
3.85a
7.37a
DMC
Expt.
12.9h
a
Farid and Needs (1992) and references therein.
Rajagopal et al. (1995).
c
Li, Ceperley, and Martin (1991).
d
Fahy, Wang, and Louie (1990a). Zero-point energy corrections of 0.18 eV for C and 0.06 eV for Si
have been added to the published values for consistency with the other data in the table.
e
Malatesta, Fahy, and Bachelet (1997).
f
Kent et al. (2000).
g
Leung et al. (1999).
h
Estimated by Knittle et al. (1989) from experimental results on hexagonal BN.
b
ties are still yielding new insights into electronic manybody phenomena. Consider a uniform electron gas with
number density n⫽(4 ␲ r s3 /3) ⫺1 . In terms of the scaled
variables r⬘i ⫽ri /r s the Hamiltonian is
Ĥ⫽⫺
1
N
兺
2r s2 i⫽1
N
2
ⵜ r⬘ ⫹
i
1
2r s i⫽1
N
兺 j⫽1
兺
(j⫽i)
1
兩 ri⬘ ⫺rj⬘ 兩
.
(5.1)
(In real calculations it is necessary to add a uniform
positive background and use a periodic Coulomb interaction as discussed in Sec. IX.C.) Note that the kineticand potential-energy operators in Eq. (5.1) scale differently as functions of r s . This means that in the veryhigh-density (small r s ) limit the interactions become
negligible and the wave function may be approximated
as a HF determinant, while at very low density (large r s )
the Coulomb interactions dominate and the kinetic energy can be ignored. In the low-density limit the electrons behave like classical charges at zero temperature
and should freeze into a so-called Wigner crystal
(Wigner, 1934). The electron gas must therefore undergo a first-order phase transition as the density is lowered.
At intermediate densities another possibility is that
the electron gas may become partially or wholly ferromagnetic. This was first predicted within the HF approximation by Bloch (1929), who considered the behavior of the HF energy as a function of the spin
polarization ␨ ⫽(N ↑ ⫺N ↓ )/N. He found that if the density was high the minimum energy was at ␨ ⫽0 and the
gas was paramagnetic, but that if r s was greater than
approximately 5.45 the minimum jumped to ␨ ⫽1 and
the gas became completely polarized. (For comparison,
Cs has an r s of 5.62.) The HF approximation is not accurate for such large values of r s and the suggestion that
the electron gas polarizes at low density has remained
controversial.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
The first attempt to study the phases of the electron
gas using DMC was made by Ceperley and Alder
(1980). Such calculations require fine comparisons of the
energies of phases with very different trial wave functions, symmetries, and finite-size errors. The energy differences of interest are so small that it is dangerous to
make precise quantitative statements, but Ceperley and
Alder did indeed observe the expected paramagnetic →
ferromagnetic → Wigner crystal sequence as the density
was lowered. They found that the ␨ ⫽1 ferromagnetic
gas was stable for r s values from 75 to 100, at which
point the bcc Wigner crystal became more favorable.
The possibility of partial polarization was considered by
Alder, Ceperley, and Pollock (1982), who found that the
50% polarized gas was more stable than the 100% polarized gas for all densities above that at which Wigner
crystallization was observed. In the region in which the
␨ ⫽1 ferromagnetic gas was more stable than both the
unpolarized gas and the Wigner crystal, the 50% polarized gas was therefore even more stable.
Very recently, Ortiz, Harris, and Ballone (1999) recalculated the energies of various phases for a range of low
to intermediate electron densities, using much larger
systems and a more careful data analysis than in earlier
work. Figure 9 shows the energy differences between the
paramagnetic and ferromagnetic gas (filled circles), and
between the ferromagnetic bcc crystal and the ferromagnetic gas (filled squares) as functions of r s . Further calculations showed that in fact the spin polarization grows
continuously from zero at r s ⫽20⫾5 to 100% at r s ⫽40
⫾5, and that the Wigner crystal is stable for r s ⬎65
⫾10. Ortiz, Harris, and Ballone (1999) also made the
controversial suggestion that the phase with partial spin
polarization may explain the recent experimental results
on Ca1⫺x Lax B6 , which exhibits very weak ferromagnetism in a narrow range of concentrations, but with
a surprisingly high Curie temperature of ⬇600 K (Cep-
Foulkes et al.: Quantum Monte Carlo simulations of solids
57
change in total energy per electron, which may be expressed in terms of ␹ and ␹ (3) as (Senatore, Moroni, and
Ceperley, 1999)
␦E
N
FIG. 9. Total energy difference (eV) times r s of 䊉, the paramagnetic and the ferromagnetic gas, and 䊏, the ferromagnetic
bcc crystal and the ferromagnetic gas. The statistical error bar
is comparable to the size of the symbols. The ferromagnetic
gas is stable when both symbols are above zero. From Ortiz,
Harris, and Ballone, 1999.
erley, 1999; Young et al., 1999).
The results of Ortiz, Harris, and Ballone (1999) are
significantly different from the results of earlier authors,
presumably because the earlier work suffered from
larger finite-size errors, although there were also other
differences in the technical details of the calculations.
Evidently there is a need for even more accurate studies
of the residual finite-size and fixed-node errors. In addition, the relative accuracies of the trial functions used
for different phases remain to be thoroughly investigated.
C. Static response of the electron gas
Many experimental techniques, including most types
of elastic and inelastic photon, neutron, and electron
scattering, measure how solids respond to small perturbations. The results of such experiments are dynamical
response functions, the most familiar of which is the dynamical susceptibility ␹ (r,r⬘ ,t⫺t ⬘ ). Although timedependent response functions such as ␹ (r,r⬘ ,t⫺t ⬘ ) are
difficult to obtain using QMC techniques (see Sec.
VI.C), both the linear and nonlinear responses to static
perturbations are straightforward to calculate. If, for example, one applies a time-independent external potential ␾ ext(r)⫽2 ␾ q cos(q•r) to a uniform electron gas with
electronic charge density ␳ 0 , the charge-density response at wave vector q takes the form 2 ␳ q cos(q•r),
where ␳ q may be expanded in powers of ␾ q as follows:
1
␳ q⫽ ␹ 共 q 兲 ␾ q⫹ ␹ (3) 共 q,q,⫺q兲 ␾ q3 ⫹¯ .
2
(5.2)
The prefactors ␹ and ␹ (3) (not to be confused with the
one-body ␹ term in the Jastrow factor) are known as the
linear and cubic response functions, respectively, and
may be obtained by applying potentials of several different strengths and fitting the calculated values of ␳ q to a
polynomial in ␾ q . Alternatively, one can calculate the
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
⫽
␹ 共 q 兲 2 ␹ (3) 共 q,q,⫺q兲 4
␾ ⫹
␾ q⫹¯ ,
␳0 q
4␳0
(5.3)
and fit this to a polynomial in ␾ q . Because the DMC
method can calculate energies without the need for extrapolated estimates (see Sec. III.D.3), the latter method
turns out to be the most accurate.
Calculations of this type have been carried out for
both the two-dimensional (Moroni, Ceperley, and Senatore, 1992) and three-dimensional electron gases (Bowen, Sugiyama, and Alder, 1994; Moroni, Ceperley, and
Senatore, 1995), but not yet for a real solid, where the
lack of translational symmetry complicates matters considerably. Although the static susceptibility is less interesting than the full dynamical susceptibility, it has still
been the focus of an enormous amount of theoretical
work using many different techniques (see, for example,
Singwi and Tosi, 1981). It was therefore surprising that
the QMC results exposed significant flaws in the best
known approximations. For further details see the recent review by Senatore, Moroni, and Ceperley (1999).
D. The relativistic electron gas
Calculating quantum relativistic effects for interacting
systems is a formidable task, but one simple approach is
to perturb about the nonrelativistic limit. The relativistic
correction to the interaction energy can be obtained to
order 1/c 2 by evaluating the Breit interaction within
first-order perturbation theory; see Bethe and Salpeter
(1957). Kenny et al. (1996) used this approach to calculate the relativistic correction to the exchangecorrelation energy of the homogeneous electron gas in
the density range r s ⫽0.1– 10. The expectation values of
the Breit interaction and the mass-velocity terms were
calculated with nonrelativistic wave functions using
VMC and DMC methods, and extrapolated estimation
(see Sec. III.D.3) was used to obtain more accurate expectation values.
The results enabled Kenny et al. (1996) to parametrize
the relativistic corrections to the LDA exchange and
correlation functional. Comparisons of relativistic LDA
and accurate relativistic calculations for atoms (Kenny,
Rajagopal, and Needs, 1995) revealed that the Darwin
part of the Breit interaction was well described by a
local-density approximation, but the term arising from
the retardation correction to the Coulomb interaction
was very poorly described. The results provided both
new insights and useful expressions for evaluating the
relativistic corrections that are important for heavier elements and dense plasmas.
E. Exchange and correlation energies
Section II.D sketched the main features of densityfunctional theory, which is by far the most popular
58
Foulkes et al.: Quantum Monte Carlo simulations of solids
FIG. 10. Contour plots in the (110) plane of Si in the diamond structure: (a) e VMC
(r)⫺e LDA
(r); (b) e VMC
⫺e LDA
(r). The chains
x
x
c
c
of atoms and bonds are represented schematically. The contours are in atomic units. From Hood et al., 1998 [Color].
method for calculating the electronic properties of solids. The explosion of interest in density-functional
theory during the eighties was built on the success of the
local-density approximation (LDA), which in turn was
hom
based on the accurate parametrizations of ⑀ xc
(n) published by Vosko, Wilk, and Nusair (1980) and Perdew
and Zunger (1981). Both these parametrizations were
constructed using values of the exchange-correlation energy of the uniform electron gas obtained from the pioneering DMC simulations of Ceperley and Alder (1980).
It is fair to say that these early DMC results made the
success of density-functional theory possible, and that
without them density-functional theory might never
have grown into the leviathan we know today.
The main drawback of approximate exchangecorrelation energy functionals such as the LDA and the
generalized gradient approximation (Langreth and
Mehl, 1983; Becke, 1988; Perdew et al., 1992) is that it is
very difficult to judge their accuracy or to improve them
when necessary. Most attempts to devise better approximations have been based on the coupling-constant integration formula (see, for example, Parr and Yang, 1989):
E xc 关 n 兴 ⫽
1
2
where
n̄ xc 共 r,r⬘ 兲 ⫽
冕冕
冕
1
0
n 共 r兲 n̄ xc 共 r,r⬘ 兲
dr dr⬘ ,
兩 r⫺r⬘ 兩
␭
n xc
共 r,r⬘ 兲 d␭,
(5.4)
(5.5)
␭
(r,r⬘ ) is the exchange-correlation hole of a fictiand n xc
tious system in which the strength of the electronelectron interaction has been reduced by a factor ␭ while
the external potential has been adjusted to keep the
electron density fixed at n(r). The exchange-correlation
hole describes the small ‘‘exclusion zone’’ that forms
around an electron at r due to the effects of the Pauli
principle and (for ␭⫽0) the Coulomb interaction. A
␭
sum rule guarantees that the integral of n xc
(r,r⬘ ) over r⬘
is always equal to ⫺1, independent of the position r or
the value of ␭.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Equation (5.4) suggests a natural (although by no
means unique) definition of the exchange-correlation
energy density:
e xc 共 r兲 ⫽
1
2
冕
n 共 r兲 n̄ xc 共 r,r⬘ 兲
dr⬘ .
兩 r⫺r⬘ 兩
(5.6)
One possible way to obtain this quantity is to use VMC
␭
(r,r⬘ ) for several different values of ␭
to calculate n xc
and to evaluate the integrals in Eqs. (5.5) and (5.6) numerically. This has recently been done for Si (Hood
et al., 1997, 1998) and for inhomogeneous electron gases
subject to cosine-wave external potentials (Nekovee
et al., 1999). Filippi and co-workers have calculated
exchange-correlation energy densities for a model
simple harmonic system (Filippi, Umrigar, and Taut,
1994) and for various atoms (Filippi, Gonze, and Umrigar, 1996) using a different definition of e xc that is not
based on the coupling-constant integration formula. Because they were studying small systems, these authors
were also able to calculate the exact exchangecorrelation potential. The comparisons of the QMC results with the various approximate exchange-correlation
energy functionals are instructive and may aid the development of better approximations in the future.
Figure 10 (Hood et al., 1998) shows how the exchange
and correlation components of the accurate exchangecorrelation energy density calculated using VMC differ
from those assumed in the LDA. The plot shows a cross
section in the (110) plane of crystalline silicon. It can be
seen that the LDA exchange energy density is not negative enough on average, while the LDA correlation energy density is too negative. When integrated over the
unit cell the two errors largely cancel, and thus the total
exchange-correlation energy is very accurate; this explains why LDA calculations of the properties of silicon
work so well. Nekovee et al. (1999) observed a similar
cancellation of errors in cosine-wave jellium, so this behavior seems to be fairly common.
Figure 11 (Nekovee, Foulkes, and Needs, 2000) shows
VMC
LDA
VMC
ADA
e xc ⫺e xc
and e xc
⫺e xc
for three different
59
Foulkes et al.: Quantum Monte Carlo simulations of solids
average-density than with the local-density approximation (Hood et al., 1997).
F. Compton scattering in Si and Li
FIG. 11. The upper graphs show e VMC
⫺e LDA
and e VMC
xc
xc
xc
ADA
⫺e xc
for three different strongly inhomogeneous cosinewave jellium systems. The lower graphs show the corresponding electron densities n(r) and (the negatives of the) Laplacians ⵜ 2 n(r). All three systems have the same average
electron density as a uniform electron gas with r s ⫽2. The
Fermi wave vector, Fermi wavelength, and Fermi energy of
this uniform system are denoted by k F0 , ␭ F0 , and E F0 , respectively. The applied cosine-wave potentials are of the form
V q cos(q•r) with V q ⫽2.08E F0 . The systems have different
wave vectors q as shown. From Nekovee, Foulkes, and Needs,
2000.
strongly inhomogeneous cosine-wave jellium systems.
The average-density approximation is a simple nonlocal
exchange-correlation energy functional devised by Gunnarsson, Jonson, and Lundqvist (1979). All three systems have the same average electron density as a uniform electron gas with r s ⫽2. The Fermi wave vector,
Fermi wavelength, and Fermi energy of this uniform system are denoted by k F0 , ␭ F0 , and E F0 , respectively. The
applied cosine-wave potentials are of the form
V q cos(q•r) with V q ⫽2.08E F0 , but the wave vector q is
different for each system.
The most striking feature of Fig. 11 is the strong similarity between the LDA errors and the Laplacian of the
electron density. This suggests that gradient-corrected
exchange-correlation energy functionals should include
a Laplacian dependence as well as the more familiar
gradient terms. Other authors, including Engel and
Vosko (1993), Umrigar and Gonze (1994), and Becke
(1998), have considered this possibility before, but for
different reasons. The definition of the exchangecorrelation energy density given in Eq. (5.6) is not the
one on which the generalized gradient approximation is
based (Burke, Cruz, and Lam, 1998), and so it is not
certain that the generalized gradient approximation can
be improved by the addition of Laplacian terms, but the
possibility is worth investigating. The nonlocal averagedensity approximation functional is clearly more accurate than the LDA at most points r. However, the cancellation of errors that explains the success of the LDA
in Si does not occur for the average-density approximation, and the total energy of Si is less accurate with the
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Quantum Monte Carlo methods have recently been
used to study inelastic x-ray scattering at large momentum transfers (Compton scattering) by Kralik, Delaney,
and Louie (1998) and Filippi and Ceperley (1999).
Compton scattering is a powerful technique for probing
electron correlation effects in solids and the Fermi surfaces of metals. Within the impulse approximation
(Eisenberger and Platzman, 1970) the recoiling electron
is treated as free, and the scattering cross section is proportional to the Compton profile. The Compton profile
at momentum p for direction ê is given by
J共 p 兲⫽
冕
dk n 共 k兲 ␦ 共 k•ê⫺p 兲 ,
(5.7)
where the momentum distribution n(k) can be expressed in terms of the wave function ⌿:
n 共 k兲 ⫽
N
V
冕
dr1 ,...,drN
冕
dr e ik•r
⫻⌿ * 共 r1 ,...,rN 兲 ⌿ 共 r1 ⫹r,...,rN 兲 ,
(5.8)
and there are N electrons in volume V. Kralik, Delaney,
and Louie (1998) performed VMC calculations of the
momentum density of Si, while Filippi and Ceperley
(1999) performed similar calculations for Li using both
the VMC and DMC methods.
The primary interest is in the valence electrons, although all the electrons contribute to the scattering.
Pseudopotentials were used in the QMC calculations so
that only the valence electrons were included, and corrections were added to account for the core-valence orthogonality. Experimental valence Compton profiles are
deduced by subtracting a contribution calculated for the
core electrons from the measured values. Figure 12
shows experimental valence Compton profiles for Li
(Sakurai et al., 1995) together with the valence Compton
profiles calculated by Filippi and Ceperley (1999) using
the LDA and VMC methods. The overall agreement between the shapes of the calculated and experimental valence Compton profiles is good, but both the LDA and
VMC profiles are too large at low momenta and too
small above the Fermi momentum (p F ⫽0.5905 a.u.).
Further calculations using the more accurate DMC
method confirmed these results.
Independent-particle calculations of Compton profiles
may be corrected for electron correlation effects using
the theory of Lam and Platzman (1974). This correction
is in principle exact within the framework of densityfunctional theory, although it is approximate within the
LDA. Both Kralik, Delaney, and Louie (1998) and Filippi and Ceperley (1999) found that the Lam-Platzman
correction gave a good account of the small differences
between the QMC and LDA valence Compton profiles.
This indicates that the approximate treatment of the valence electron correlations is not the primary reason for
60
Foulkes et al.: Quantum Monte Carlo simulations of solids
alike. High pressures are simple to achieve in calculations, but quantum effects due to the small mass of the
proton become increasingly important. In addition, the
energy differences between candidate structural phases
are very small and predictions from mean-field calculations may not be trustworthy.
Quantum Monte Carlo techniques are ideal for studying the hydrogen problem because they allow an accurate treatment of the quantum motion of both the electrons and protons. Among the very first applications of
QMC methods to real solids was the DMC work on
solid hydrogen of Ceperley and Alder (1981, 1987),
while further DMC studies have been performed by Natoli, Martin, and Ceperley (1993, 1995). These studies
encompassed molecular and monatomic phases in which
the protons are localized around lattice sites, so that exchange effects between protons are negligible and the
wave function can be taken to be symmetric in the proton coordinates. The trial wave functions used are of the
form
冋
⌿⫽⌿ e exp ⫺
FIG. 12. Valence Compton profiles for Li in the [100], [110],
and [111] directions. Variational Monte Carlo results for the
cells with 250 and 686 atoms are compared with the LDA valence profiles from an all-electron LDA calculation and with
experiment. From Filippi and Ceperley, 1999.
the discrepancies between the calculated and measured
Compton profiles of Si and Li.
Before the advent of the QMC calculations it was
widely believed that the approximate treatment of electron correlation was responsible for the discrepancies
between LDA and measured Compton profiles. These
QMC studies will force a more serious consideration of
other possibilities such as thermal effects and the possible inadequacy of the impulse approximation for calculating the contribution from the core electrons.
G. Solid hydrogen
Hydrogen, the most abundant of all elements, has
been the focus of a great deal of experimental and theoretical research. One interesting area of application is
to the interiors of giant planets. For example, Jupiter is
about 90% hydrogen at high pressures and temperatures, most of which exists in a fluid metallic state. At
low temperatures and pressures hydrogen solidifies into
the only known example of a molecular quantum crystal,
in which the molecules rotate freely due to quantum
rather than thermal effects. With sufficient applied pressure the molecules lock into preferred orientations, and
at very high pressures hydrogen is expected to form a
metallic solid, which could exhibit interesting properties
such as high-temperature superconductivity. Metallization, defined as the existence of a finite dc electrical conductivity at low temperatures, has not, however, been
observed in terrestrial static pressure experiments on hydrogen. Understanding the behavior of hydrogen at high
pressures is a challenge to experimentalists and theorists
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
兺
␣⫽␤
u pp共 d␣ ⫺d␤ 兲 ⫺
册
兺␣ c ␣兩 d␣ ⫺d̄␣兩 2 ,
(5.9)
where d␣ and d̄␣ denote, respectively, the proton positions and the lattice sites. The factor ⌿ e is the electronic
part of the wave function for fixed proton positions and
is taken to be a standard Slater-Jastrow form, while the
symmetric proton part consists of a proton-proton correlation term u pp and a product of Gaussian functions
that localize the protons around the lattice sites. Ceperley and Alder (1987) explored both molecular and
atomic phases, predicting a transition from the freely
rotating molecular quantum crystal to a rotationally ordered molecular phase at about 100 GPa, which was
subsequently identified in experiments (Lorenzana, Silvera, and Goettel, 1990), and a further transition to
atomic phases at about 300 GPa. The DMC calculations
of Natoli, Martin, and Ceperley (1993) showed that the
zero-point motion of the protons is a very important
factor in determining the relative stabilities of the highpressure monatomic phases of hydrogen. The electronic
energy favors low coordination numbers at these densities, but this is counteracted by the effect of the zeropoint energy, which favors high coordination and symmetry. At still higher pressures the Madelung energy of
the proton lattice should result in a transition to closepacked structures. Natoli, Martin, and Ceperley (1993)
found the diamond structure to be the most stable of the
monatomic phases up to pressures of about 400 GPa.
They concluded that the usual estimates of zero-point
energies from mean-field theories assuming the harmonic approximation were inaccurate, in some cases by
a factor of two. Natoli, Martin, and Ceperley (1995)
found that insulating molecular phases with canted orientations of molecules on an hcp lattice were favored at
pressures above about 123 GPa. Combining these data
with their earlier DMC results on monatomic phases of
hydrogen, they predicted that the transition to a monatomic diamond-structure phase would occur at around
Foulkes et al.: Quantum Monte Carlo simulations of solids
61
FIG. 13. The equation of state of hydrogen determined by
DMC calculations and compared with extrapolations from experiment due to Hemley et al. (1990). Diffusion Monte Carlo
data are shown for the low-pressure Pa3 molecular structure
(Ceperley and Alder, 1987), the diamond structure (Natoli,
Martin, and Ceperley, 1995), and two molecular phases consisting of differing orientations of molecules on an hcp lattice
(mhcp-o and mhcp-c) (Natoli, Martin, and Ceperley, 1993).
From Natoli, Martin, and Ceperley, 1995.
300 GPa (see Fig. 13). Such a transition pressure would
be consistent with values estimated by extrapolating the
measured frequency of the molecular stretching mode to
zero.
Both experimental and theoretical results suggest that
the metallization of hydrogen occurs at a pressure not
too far above the largest static pressure currently attainable of about 300 GPa. The unique capabilities of the
DMC method have already made substantial contributions to the theoretical understanding of solid hydrogen
at zero temperature. The path-integral Monte Carlo
method has been applied to hydrogen at high temperatures (Pierleoni et al., 1994; Magro et al., 1996; Militzer,
Magro, and Ceperley, 1999), and it is hoped that further
DMC and path-integral Monte Carlo studies will help in
obtaining a more complete picture of the phase diagram
of hydrogen.
FIG. 14. Geometries and charge densities of C20 isomers calculated using the HF method. From Grossman, Mitas, and
Raghavachari, 1995 [Color].
ating the energetic ordering of three isomers (ring,
graphitic bowl, dodecahedron cage; see Fig. 14) of C20 in
an attempt to search for the smallest carbon fullerene
(Grossman, Mitas, and Raghavachari, 1995). Raghavachari and co-workers (1993) had earlier discovered large
discrepancies in the energy differences between the
three isomers as calculated using the LDA and generalized gradient approximation methods. For example, the
LDA predicts the cage structure to be about 4 eV lower
in energy than the ring, while the Becke-Lee-Yang-Parr
(BLYP) generalized gradient approximation functional
predicts essentially the opposite (see Fig. 15). Such dif-
H. Clusters
Clusters of atoms are intermediate between small
molecules and solids and exhibit a rich variety of physical and chemical properties. They also form an important testing ground for QMC methods because of the
absence of the finite-size errors that plague periodic
boundary conditions calculations. This facilitates comparisons between QMC and other computational methods as well as comparisons with experimental data.
Encouraging progress in applying QMC methods was
demonstrated in studies of silicon and carbon clusters of
intermediate sizes up to 20 atoms (Grossman and Mitas,
1995; Grossman, Mitas, and Raghavachari, 1995; Mitas
and Grossman, 1997). These studies focused on the
binding (atomization) energies and the energetic ordering of competing isomers. Perhaps the most striking result was obtained by Grossman and co-workers in evaluRev. Mod. Phys., Vol. 73, No. 1, January 2001
FIG. 15. Relative energies of C20 isomers from the HF, LDA,
BLYP, and DMC methods. The energies are given relative to
the lowest-energy isomer within the given theory. From Grossman, Mitas, and Raghavachari, 1995.
62
Foulkes et al.: Quantum Monte Carlo simulations of solids
ferences are unusually large for moderate-sized systems.
Diffusion Monte Carlo calculations with HF geometries
showed that the graphitic bowl was the lowest-energy
isomer. The results in Fig. 15 were met with surprise and
even disbelief in the electronic structure community,
and so it is reassuring that an independent method confirmed the DMC results. Calculations by Murphy and
Friesner (1998), using their newly developed high-level
perturbation approach based on generalized valencebond wave functions, agreed with the DMC results to
within the error bars. The QMC calculations on C20
therefore provided new and unexpected results that
would have been very difficult to obtain by any other
method, and highlighted the limited predictive power of
LDA/generalized gradient approximation approaches
for systems that were thought to be well understood.
The study of carbon clusters within DMC was further
extended to C24 , C26 , C28 , and C32 by Kent et al. (2000).
These authors found that the cage structure of C24 is
higher in energy than other isomers, but that for C26 and
C28 cage structures are slightly favored over other isomers. For C32 the cage geometry is very clearly the most
stable.
I. Formation energies of silicon self-interstitials
Silicon is the most important material in the microelectronics industry, and the diffusion of dopant impurity atoms during thermal processing is one of the factors that limits how small semiconductor devices can be
made. The diffusion of impurity atoms in silicon is critically influenced by intrinsic defects such as selfinterstitials and vacancies, and it is therefore of great
importance to improve our understanding of these defects. Unfortunately it has not been possible to detect
self-interstitials directly, although their presence has
been inferred using various techniques (Fahey, Griffin,
and Plummer, 1989). Measurements of the selfdiffusivity D SD of silicon at high temperatures, using radioactive isotopes of silicon as tracers, have established
an Arrhenius behavior with an activation energy in the
range of 4.1–5.1 eV (Frank et al., 1985). D SD is usually
written as the sum of contributions from independent
diffusive mechanisms, and these contributions can each
be written as the product of the diffusivity D i and the
concentration C i of the relevant defect, i.e.,
D SD⫽
兺i D i C i .
(5.10)
It is widely believed that self-interstitial diffusion is important at higher temperatures, while vacancy diffusion
is important at lower temperatures (see the review by
Gösele, Plössl, and Tan, 1996). The experimental situation regarding self-diffusion in silicon is, however, still
highly controversial, especially when it comes to the individual values of D i and C i . Indeed, experimental data
have been used to support values of the diffusivity of the
silicon self-interstitial that differ by ten orders of magniRev. Mod. Phys., Vol. 73, No. 1, January 2001
tude at the temperatures of around 800 °C at which silicon is processed (Eaglesham, 1995).
Many theoretical studies of self-interstitials in silicon
have been carried out. The most advanced of these have
used LDA-density-functional methods to calculate the
defect formation energies and energy barriers to diffusion (Bar-Yam and Joannopoulos, 1984; Blöchl et al.,
1993). The consensus view emerging from these calculations is that the split-具 110典 , hexagonal, and tetrahedral
self-interstitial defects are the lowest in energy. Another
interesting suggestion is that self-diffusion could occur
without point defects via exchange of neighboring atoms
in the perfect lattice, and Pandey (1986) has proposed
such a concerted exchange mechanism for self-diffusion
in silicon. The structures of the split-具 110典 , hexagonal,
and tetrahedral interstitial defects and of the saddle
point of Pandey’s concerted-exchange mechanism are illustrated in Fig. 16.
Leung et al. (1999) performed fixed-node DMC calculations and density-functional calculations using the
LDA and the PW91 generalized gradient approximation
to determine the formation energies of self-interstitials
in silicon. Within each method they found the split-具 110典
and hexagonal interstitials to be the most stable, although the formation energies were significantly different in the three methods (see Table II). The DMC formation energies are about 1 eV larger than the PW91
generalized gradient approximation values and 1.5 eV
larger than the LDA values. Leung et al. (1999) used
these DMC data to estimate a value for the activation
energy for self-interstitial diffusion of about 5 eV, which
is consistent with the value deduced from experiment of
4.84 eV (Gösele, Plössl, and Tan, 1996). The activation
energies predicted by the LDA and PW91 generalized
gradient approximation density functionals are, however, considerably lower than the experimental value
and do not provide a satisfactory explanation of selfdiffusion in silicon. This study has highlighted the importance of a proper treatment of electron correlation when
treating such systems.
J. Jellium surfaces
The jellium model with the positive background terminated at a plane provides the simplest model of a
metal surface. Many theoretical techniques have been
used to investigate jellium surfaces, which have become
a testing ground for developing new methods of studying correlation effects in inhomogeneous systems. Li
et al. (1992) used DMC to study the surface of jellium at
a density of r s ⫽2.07, which is the average valence
charge density of aluminum. Acioli and Ceperley (1996)
considered five densities in the range r s ⫽1.87– 3.93, calculating the surface energies, work functions, charge
densities, and pair-correlation functions. The trial wave
functions were of the Slater-Jastrow type with a determinant of LDA orbitals. The final QMC charge densities
agreed with the LDA densities to within 2%, supporting
the use of LDA orbitals for the trial wave function. The
DMC surface energies were higher than the LDA and
63
Foulkes et al.: Quantum Monte Carlo simulations of solids
FIG. 16. Self-interstitial defects in silicon: (a) the split 具110典, (b) hexagonal, and (c) tetrahedral interstitial defects, and (d) the
saddle point of the concerted–exchange mechanism. The atom(s) forming the defect are shown in red, while the nearest neighbors
to the defect atoms are shown in yellow. The bonds between the defect and nearest-neighbor atoms are shown in orange. From
Leung et al., 1999 [Color].
TABLE II. Local-density approximation (LDA), PW91 generalized gradient approximation (GGA),
and diffusion Monte Carlo (DMC) formation energies in eV of the self-interstitial defects and the
saddle point of the concerted-exchange mechanism. Note that the DMC results for the 16- and
54-atom simulation cells are consistent, indicating that the residual finite-size effects are small.
Defect
Split-具110典
Hexagonal
Tetrahedral
Concerted exchange
a
16-atom supercell.
54-atom supercell.
b
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
LDA
GGA
DMCa
DMCb
3.31
3.31
3.43
4.45
3.84
3.80
4.07
4.80
4.96(24)
4.70(24)
5.50(24)
5.85(23)
4.96(28)
4.82(28)
5.40(28)
5.78(27)
64
Foulkes et al.: Quantum Monte Carlo simulations of solids
generalized gradient approximation energies, although
quite close to the Fermi-hypernetted-chain (FHNC) results of Krotscheck, Kohn, and Qian (1985) at low densities. The work functions were in reasonable agreement
with the FHNC results and were lower than the LDA
values. Inside the jellium the pair-correlation functions
were nearly spherical, but as the electron moved towards the surface they showed significant anisotropy.
The large discrepancy between the DMC and densityfunctional surface energies is surprising and has recently
been questioned by Yan et al. (2000). They observe that
the density-functional and DMC energies (Ballone, Umrigar, and Delaly, 1992) of finite jellium spheres are in
rather close agreement and suggest that the discrepancy
in the case of the surface may arise from finite-size errors in the DMC simulations.
VI. EXCITED STATES
A. Introduction
Although the VMC and DMC methods were designed
to study ground states, they can also provide some information about excited states. The lowest band gap of a
solid, which may be measured from a combination of
photoemission and inverse photoemission experiments,
is the difference between the energies to add and subtract an electron from the N-electron system: E g
⫽(E N⫹1 ⫺E N )⫺(E N ⫺E N⫺1 ). This expression involves
only ground-state energies and hence is immediately accessible to QMC methods. The first application of DMC
to an energy gap in a solid was by Ceperley and Alder
(1987), who calculated the minimum energy zone center
gap of the Pa3 molecular hydrogen crystal as a function
of pressure. To evaluate the energies of the N⫹1 and
N⫺1 electron systems they added uniform neutralizing
background charge densities to preserve the charge neutrality of the simulation cell.
Another straightforward way of obtaining excitation
energies is to devise a many-electron trial wave function
that models an excited state and use it in a VMC calculation. If the chosen trial wave function has a specific
symmetry, the variational principle guarantees that the
energy obtained is greater than or equal to the eigenvalue of the lowest exact eigenstate of that symmetry.
At first sight it might appear that the DMC method is
inapplicable to excited states because the wave function
always evolves towards the ground state. In fixed-node
DMC, however, the nodal constraint ensures convergence to the lowest energy state compatible with the
imposed nodal surface, not to the overall ground state. It
is straightforward to show that, if the nodal surface of
the trial wave function is the same as that of an exact
eigenstate, then the fixed-node DMC algorithm gives
the exact energy of that eigenstate.
Until recently, it was widely believed that a symmetryconstrained variational theorem analogous to the one
for VMC also held for fixed-node DMC calculations.
It was assumed that the symmetry of the fixed-node
ground state produced by the imaginary-time evolution
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
was always the same as that of the trial function used to
define the nodes, and hence that the DMC energy was
always greater than or equal to the eigenvalue of the
lowest exact eigenstate of that symmetry. It has now
been shown (Foulkes, Hood, and Needs, 1999) that this
is not necessarily the case unless the symmetry of interest corresponds to a one-dimensional irreducible
representation of the symmetry group of the Hamiltonian. In other words, the eigenstate should be nondegenerate or have only accidental degeneracies. For multidimensional irreducible representations the symmetry
of the fixed-node ground state may differ from that of
the trial state, and hence the DMC energy may lie below
the eigenvalue of the lowest exact eigenstate with the
same symmetry as the trial state. In such cases it is possible to obtain weaker variational principles by choosing
trial functions that transform according to onedimensional irreducible representations of subgroups of
the full symmetry group (Foulkes, Hood, and Needs,
1999). In practice, the surprise is that the fixed-node
DMC method often works well for excited states regardless of the dimension of the irreducible representation; it
even works quite well for excited states that are not the
lowest energy states of any particular symmetry.
B. VMC and DMC calculations of excitations in solids
Variational and diffusional Monte Carlo calculations
of excitations in solids are computationally demanding
because excitation energies are ‘‘1/N’’ effects; that is, the
fractional change in the total energy due to the presence
of the excitation is inversely proportional to the number
of electrons in the simulation cell. Excited-state VMC
and DMC simulations therefore require great statistical
accuracy and a careful treatment of finite-size effects.
Mitas and Martin (1994) used the DMC method to
estimate the band gap of atomic solid nitrogen in the
I2 1 3 structure. Simulations were carried out to obtain
the total energy of the ground state and of an excited
state in which a single electron had been promoted from
the highest occupied ⌫ state to the lowest unoccupied H
state. This can be interpreted as a calculation of the energy of an electron-hole pair (Mott-Wannier exciton).
The exciton binding energy is usually small (⬃0.01–0.1
eV) and so it is easy to obtain the excitation energy once
the formation energy of the exciton is known.
Mitas (1996) later used the DMC method to calculate
a selection of excitation energies in diamond, and a
more complete study has recently been completed by
Towler, Hood, and Needs (2000); see Table III. The
DMC approach gives a good estimate of the ⌫ 25⬘ →X 1c
band gap but somewhat overestimates the width of the
valence band.
Williamson et al. (1998) used the DMC method to calculate a large number of excitation energies in Si. By
studying many different electron-hole pair excitations,
Williamson et al. obtained the quasiparticle energies
shown in Fig. 17. The solid lines show an empirical
pseudopotential band structure that may be taken as an
Foulkes et al.: Quantum Monte Carlo simulations of solids
65
TABLE III. Excitation energies (eV) of diamond calculated using the HF, LDA, GW, and DMC
methods, and compared with experimental data.
Method
HF
LDA
GW
DMC
Expt.
Band gap
⌫ 25⬘ →X 1c
Bandwidth
⌫ 1 v →⌫ 25⬘
13.2a
4.6,a 4.63b
6.3c
6.0(4),a 5.71(20)b
6.1f
29.4a
22.1,a 21.35b
22.88,c 23.0d
23.9(7),a 24.98(20)b
23.0(2),e
a
Mitas (1996).
Towler, Hood, and Needs (2000).
c
Rohlfing et al. (1993).
d
Hybertsen and Louie (1986).
e
Jiménez et al. (1997).
f
Estimated by correcting the measured minimum band gap of 5.48 eV.
b
accurate representation of the experimental results. The
DMC energy of the state at the top of the valence band
is set equal to zero by definition, but the rest of the
results are meaningful. The low-lying quasiparticle energies are accurate, but the energies of holes lying deeper
in the valence bands are significantly overestimated due
to deficiencies of the trial wave functions. A new feature
of this work was the successful calculation of several
different excited states at the same k point. Moreover, it
was found that the DMC method produced equally good
results whether or not the electron and hole had the
same crystal momentum.
with the VMC algorithm to calculate the band structure
of Si and obtained results not much worse than those
found using the direct DMC method. The main advantage of the extended Koopmans’ theorem approach is
that it allows many quasiparticle energies to be calculated simultaneously. Within the so-called diagonal approximation, which is accurate in Si (Kent et al., 1998),
the extended Koopmans’ theorem reduces to the
scheme used previously by Fahy, Wang, and Louie
(1990b) to calculate hole energies in Si, and by Tanaka
(1995) to calculate hole energies in NiO.
VII. WAVE-FUNCTION OPTIMIZATION
C. Other QMC methods for excited states
A. Introduction
There are a number of alternative methods for calculating excitation energies within QMC. Ceperley and
Bernu (1988) combined the idea of the generalized
variational principle (i.e., the variational principle for
the energies of a set of orthogonal trial functions) with
the DMC algorithm to derive a method for calculating
the eigenvalues of several different excited states simultaneously. The first application of the Ceperley-Bernu
method was to vibrational excited states (Bernu, Ceperley, and Lester, 1990), but it has also been used to investigate electronic excitations of the two-dimensional uniform electron gas (Kwon, Ceperley, and Martin, 1996)
and of He atoms in strong magnetic fields (Jones, Ortiz,
and Ceperley, 1997). Correlated sampling techniques
(Kwon, Ceperley, and Martin, 1996; Jones, Ortiz, and
Ceperley, 1997) can be used to reduce the variance, but
the Ceperley-Bernu method has stability problems in
large systems and has not been applied to a real solid.
Another method that uses the generalized variational
principle is based on the extended Koopmans’ theorem
derived independently by Day, Smith, and Garrod
(1974) and Morrell, Parr, and Levy (1975). The extended Koopmans’ theorem leads to an approximate expression for the ground- and excited-state energies of
the N⫹1 and N⫺1 electron systems relative to the
ground-state energy of the N electron system. Recently,
Kent et al. (1998) used this expression in conjunction
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
The quality of the trial wave function controls the statistical efficiency of the VMC and DMC algorithms and
determines the final accuracy obtained. Clearly one
would like to use a high-quality trial wave function, but
there is also an issue of computational efficiency. The
most costly part of VMC and DMC calculations is nor-
FIG. 17. The DMC band structure of Si (Williamson et al.,
1998). The solid lines show the empirical pseudopotential band
structure of Chelikowsky and Cohen (1976).
66
Foulkes et al.: Quantum Monte Carlo simulations of solids
mally the repeated evaluation of the trial wave function
(and its gradient and Laplacian). It is therefore important to use a trial wave function that is as accurate as
possible yet can be computed rapidly.
By far the most common type of trial wave function
used in VMC and DMC calculations for atoms, molecules, and solids is the Slater-Jastrow form discussed in
Sec. IV. Ideally one would like to perform a simultaneous optimization of the u and ␹ functions in the Jastrow factor and the orbitals in the Slater determinants.
However, almost all work on large systems has so far
involved optimizing only the u and ␹ functions. Typical
solid-state problems currently involve optimizing of order 102 parameters for 102 – 103 electrons.
B. The cost function
To optimize a wave function containing a set of parameters 兵 ␣ 其 requires some criterion for deciding on the
quality of a particular parameter set. To this end we
require a cost function, which is to be minimized with
respect to the values of the parameters. The choice of
cost function may depend on the application. For instance, if one wants to calculate the best variational
bound on the energy in a VMC calculation one should
minimize the variational energy,
E V共 ␣ 兲 ⫽
冕
⌿ T2 共 ␣ 兲 E L 共 ␣ 兲 dR
冕
,
(7.1)
冕
冕
,
(7.2)
⌿ T2 共 ␣ 兲 dR
⌿ T 共 ␣ 兲 ⌽ 共 t→⬁ 兲 dR
冕
冕
⌿ T2 共 ␣ 0 兲 w 共 ␣ 兲关 E L 共 ␣ 兲 ⫺E V 共 ␣ 兲兴 2 dR
冕
.
(7.3)
⌿ T2 共 ␣ 兲 dR
Maximization of the overlap (Ceperley and Alder, 1987)
has rarely been used because it requires a computationally intensive DMC calculation at each step of the optimization.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
,
⌿ T2 共 ␣ 0 兲 w 共 ␣ 兲
dR
(7.4)
where
E V共 ␣ 兲 ⫽
⌿ T2 共 ␣ 兲关 E L 共 ␣ 兲 ⫺E V 共 ␣ 兲兴 2 dR
which minimizes the statistical error bar on the VMC
energy. Another option is to minimize some combina2
2
/E V
tion of the variance and energy, such as ␴ E
(Meierovich, Mushinski and Nightingale, 1996). To
maximize the accuracy of an extrapolated estimator (see
III.D.3) one wants the trial and DMC wave functions to
be as close as possible, which can be achieved by maximizing their overlap,
冕
␴ E2 共 ␣ 兲 ⫽
⌿ T2 共 ␣ 兲 dR
where E L ⫽⌿ T⫺1 ( ␣ )Ĥ⌿ T ( ␣ ) is the local energy. It has
also been suggested that minimizing the energy will
maximize the efficiency of a DMC calculation (Ceperley, 1986). An alternative is to minimize the variance of
the energy,
␴ E2 共 ␣ 兲 ⫽
In practice almost all wave-function optimizations
have been performed by minimizing the variance of the
energy. A number of reasons have been advanced for
preferring variance minimization to energy minimization, but the most important is that it is more stable in
large systems, as explained below. Optimizing wave
functions by minimizing the variance of the energy is
actually a very old idea, having been used in the 1930s.
The first application using Monte Carlo techniques to
evaluate the integrals appears to have been by Conroy
(1964), but the present popularity of the method derives
from the work of Umrigar and collaborators (Umrigar,
Wilson, and Wilkins, 1988).
Let us look at the variance of the energy and the minimization procedure in more detail. The minimum pos2
sible value of ␴ E
is zero, which is obtained if and only if
⌿ T ( ␣ ) is an exact eigenstate of Ĥ. The variance therefore has a minimum at every eigenstate of Ĥ, which
makes it a suitable cost function for optimizing both
ground and excited states, although it has not been ap2
plied widely to excited states. Minimization of ␴ E
is normally carried out via a correlated-sampling approach in
which a set of configurations distributed according to
⌿ T2 ( ␣ 0 ) is generated, where ␣ 0 is an initial set of param2
( ␣ ) is then evaluated as
eter values. The variance ␴ E
冕
⌿ T2 共 ␣ 0 兲 w 共 ␣ 兲 E L 共 ␣ 兲 dR
冕
,
⌿ T2 共 ␣ 0 兲 w 共 ␣ 兲
(7.5)
dR
and the integrals contain a weighting factor w( ␣ ) given
by
w共 ␣ 兲⫽
⌿ T2 共 ␣ 兲
⌿ T2 共 ␣ 0 兲
.
(7.6)
2
The parameters 兵 ␣ 其 are adjusted until ␴ E
( ␣ ) is minimized.
The advantage of the correlated-sampling approach is
that one does not have to generate a new set of configurations every time the parameter values are changed. In
practice the set of configurations is normally regenerated a few times (typically three or four) during the optimization procedure because the estimate of the variance is poor if 兩 兵 ␣ 0 其 ⫺ 兵 ␣ 其 兩 is too large. A variant of this
scheme is to replace the energy E V ( ␣ ) by a fixed value
Ē that is a little below the ground-state energy. Minimization of this modified cost function is equivalent to
2
minimizing a linear combination of E V and ␴ E
.
Foulkes et al.: Quantum Monte Carlo simulations of solids
C. Numerical stability of variance minimization
The cost function is evaluated as an average over a set
of M configurations Rm drawn from the distribution
⌿ T2 ( ␣ 0 ),
M
␴ E2 共 ␣ 兲 ⯝
兺m w 共 Rm ; ␣ 兲关 E L共 Rm ; ␣ 兲 ⫺E V共 兵 Rm 其 ; ␣ 兲兴 2
M
.
兺m w 共 Rm ; ␣ 兲
(7.7)
2
The eigenstates of Ĥ give the minimum value, ␴ E
⫽0,
for any set of configurations because E L (R) is independent of R for an eigenstate (Nightingale and Umrigar,
1997; Kent, Needs, and Rajagopal, 1999). This highly
desirable feature is not shared by the energy E V . The
fact that the positions of the global minima of the variance are robust to finite sampling is an important advantage of variance minimization over energy minimization.
Direct minimization of the variance of Eq. (7.7) has
often been successful. However, in large systems the
procedure often exhibits a numerical instability: as the
minimization proceeds, a few configurations (often only
one) acquire a very large weight. The estimate of the
variance is then reduced almost to zero by a set of parameters that usually give extremely poor results in subsequent QMC calculations. This instability has been noticed by many researchers and also occurs in energy
minimization. In principle one could overcome it by using more configurations, but the number required is normally impossibly large. Fortunately, in variance minimization (though not in energy minimization) there is a
simple remedy, which relies on the observation that the
positions of the minima of the variance are not affected
by the values of the weights as long as they are positive
(Kent, Needs, and Rajagopal, 1999). This is because the
exact variance of Eq. (7.2) reaches its minimum value of
zero if and only if the trial function is an exact eigenfunction, in which case the local energy is independent
of position. The numerical instability may therefore be
eliminated by restricting the values of the weights. In
some calculations the upper value of the weight is limited (Filippi and Umrigar, 1996), while in others the
weights are set to unity (Schmidt and Moskowitz, 1990;
Williamson et al., 1996). This freedom is not available in
energy minimization because altering the weights alters
the positions of the minima.
D. Minimization procedures
The minimization itself may be carried out using standard optimization techniques such as the LevenbergMarquardt method (Press et al., 1992), which finds the
unconstrained minimum of a sum of squares and requires only the function values. If one is optimizing only
the Jastrow factor then the storage and CPU requirements may be greatly reduced by using u and ␹ functions that are linear in the parameters 兵 ␣ 其 (Williamson
et al., 1996). This allows the sums over the electron coRev. Mod. Phys., Vol. 73, No. 1, January 2001
67
ordinates to be calculated once for each set of configurations, instead of every time the parameter values are
changed. An alternative method for performing the optimization is to use the stochastic gradient approximation, which was introduced into QMC calculations by
Harju et al. (1997). The stochastic gradient approximation has the appealing feature that it was designed for
optimizations in the presence of noise. However, our
experience is that deterministic methods of minimization over a fixed set of configurations perform much better than the stochastic gradient approximation.
The single-particle orbitals in the Slater determinants
are normally obtained from mean-field calculations and
will not be optimal in the presence of the Jastrow factor.
Direct optimization of the single-particle orbitals using
variance minimization is possible in small systems (Umrigar, Nightingale, and Runge, 1993) but would be very
expensive in large systems because the number of parameters increases rapidly with system size. A more
promising technique (Fahy, 1999; Filippi and Fahy,
2000) is to optimize the potential that generates the orbitals rather than the orbitals themselves. The equation
determining the orbitals is obtained by minimizing the
fluctuations in the local energy. This method corresponds to energy minimization and is limited in the
sense that the orbitals are constrained to be the lowestenergy eigenstates of a one-electron Hamiltonian. However, it has the significant advantage that the number of
parameters in the potential does not increase as rapidly
with system size as the number of parameters in the
single-particle orbitals.
VIII. PSEUDOPOTENTIALS
A. The need for pseudopotentials
Quantum Monte Carlo methods have been applied
very successfully to first-row atoms, but the computational effort increases rapidly with the atomic number
Z. Various estimates of the scaling of the computational
cost with Z have been made, with results ranging from
Z 5.5 (Ceperley, 1986) to Z 6.5 (Hammond, Reynolds, and
Lester, 1987). This scaling effectively rules out applications to heavy atoms. The presence of core electrons
causes two related problems. Because of the shorter
length scales associated with variations in the wave function near a nucleus of large Z, the time step should be
decreased. This problem can be significantly reduced by
the use of acceleration schemes, such as those devised
by Umrigar (1993) and Stedman, Foulkes, and Nekovee
(1998). The other problem is that the fluctuations in the
local energy tend to be large near the nucleus because
both the kinetic and potential energies are large. Although these fluctuations can be reduced by a judicious
choice of trial wave function, in practice they are large
for heavier atoms.
However, many properties of interest, including the
interatomic bonding and the low-energy excitations, are
determined by the behavior of the valence electrons,
and just as in other electronic structure techniques one
68
Foulkes et al.: Quantum Monte Carlo simulations of solids
can use pseudopotentials to remove the core electrons
from the problem. This approach serves to reduce the
effective value of Z. Errors are inevitably introduced,
but the gain in computational efficiency is very large and
is sufficient to make applications to heavy atoms feasible.
B. Nonlocal pseudopotentials
The ideas behind pseudopotentials are perhaps best
explained in the context of independent-electron
schemes such as the Hartree-Fock method and
Hohenberg-Kohn-Sham density-functional theory. In
both these theories the wave function consists of a determinant of orbitals, which can be partitioned into core
and valence orbitals on chemical grounds. For example,
the ground-state configuration of the silicon atom is
关 1s 2 2s 2 2p 6 3s 2 3p 2 ]. The 3s and 3p orbitals take part in
the chemical bonding and are designated as valence orbitals, while the 1s, 2s, and 2p orbitals largely retain
their atomic identities and are designated as core orbitals. The idea is to create an effective potential (the
pseudopotential) that reproduces the effects of both the
nucleus and the core electrons on the valence electrons.
This is done separately for each of the different angular
momentum states, so the pseudopotential contains angular momentum projectors and is therefore a nonlocal
operator. For silicon the lowest-energy states of the
pseudopotential with angular momenta l⫽0,1,2 are the
pseudo-3s, 3p, and 3d orbitals, respectively.
It is conventional to divide the pseudopotential V ps
l (r)
ps
for any given atom into a local part V loc
(r) common to
ps
(r), for anall angular momenta and a correction, V nl,l
gular momentum l. The final result is an effective interaction potential between a valence electron at r and the
(pseudo) ion at the origin. The electron-ion potentialenergy term in the full many-electron Hamiltonian of
the atom then takes the form
V loc共 R兲 ⫹V̂ nl⫽
ps
ps
,
共 r i 兲 ⫹ 兺 V̂ nl,i
兺i V loc
i
(8.1)
ps
where V̂ nl,i
is a nonlocal operator that acts on an arbitrary function of ri as follows:
ps
V̂ nl,i
f 共 ri 兲 ⫽
ps
* 共 ⍀ i⬘ 兲 f 共 ri⬘ 兲 d⍀ i⬘ .
V nl,l
共 r i 兲 Y lm 共 ⍀ i 兲 冕 Y lm
兺
l,m
4␲
(8.2)
The angular integrals pick out the different angular momentum components (s,p,d,...) of the function f(ri )
and so guarantee that each symmetry channel
ps
(s,p,d,...) ‘‘feels’’ its own potential V nl,l
(r).
The first norm-conserving LDA pseudopotentials
were generated by Hamann, Schlüter, and Chiang
(1979). In a similar development, Christiansen, Lee, and
Pitzer (1979) generated HF pseudopotentials, which are
called effective core potentials. Both of these sets of
pseudopotentials have been used very successfully in numerous applications, including QMC simulations.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Within independent-electron theories such as the
density-functional and HF theories, the distinction between core and valence electrons (the ‘‘core-valence
partition’’) is exact and is achieved by simply designating some orbitals as core orbitals and others as valence
orbitals. In a many-electron framework, however, the
electrons are identical particles and the core-valence
partition cannot be exact. The use of pseudopotentials in
many-electron calculations therefore involves a further
approximation. The accurate results obtained in pseudopotential QMC simulations show that this extra approximation is usually of minor importance.
C. Core-polarization potentials
Correlation between electrons within a core has an
indirect effect upon the valence electrons, which is usually unimportant. The neglect of core-valence correlation can, however, be significant whenever the number
of valence electrons is small and the core is large; the
core then exhibits core-polarization and relaxation effects as a response to changes in the valence environment. Examples of systems with large core-polarization
effects include alkali metals with one or two valence
electrons.
One can include some of the effects of core-valence
correlation within a pseudopotential framework by introducing a core-polarization potential. This can be obtained within a core-valence partition scheme using the
theory of Callaway (1957), although in practice empirical forms have been employed. Callaway (1957) showed
that, after a series of approximations, the corepolarization potential could be approximated by a dipole term, which represents the polarization of the core
due to the electric field of the valence electrons (and
other ions, if present). It is consistent to use a corepolarization potential in conjunction with a HF pseudopotential.
Rather than calculate the core-polarization potential
from first principles, it has normally been described by a
simple analytic form. Core-polarization potentials have
been generated for a wide range of elements using several different methods (Müller, Flesch, and Meyer, 1984;
Müller and Meyer, 1984; Shirley and Martin, 1993), and
it has been shown (Shirley and Martin, 1993) that the
combination of HF pseudopotentials and corepolarization potentials works very well for the first two
rows of the periodic table. The use of core-polarization
potentials within QMC methods was tested in the case
of the sodium dimer by Shirley, Mitas, and Martin
(1991), and the results were significantly better than
those obtained using pseudopotentials that neglected
core-polarization effects.
D. Pseudopotentials in variational Monte Carlo
In this subsection we describe the technical aspects of
using nonlocal pseudopotentials in VMC. The first
pseudopotential QMC calculations made the approximation that the nonlocal angular momentum projection
69
Foulkes et al.: Quantum Monte Carlo simulations of solids
operators acted only on the determinantal part of the
wave function (Hammond, Reynolds, and Lester, 1987;
Hurley and Christiansen, 1987; Christiansen, 1988).
Fahy, Wang, and Louie (1988, 1990a) formulated and
applied the VMC method with nonlocal pseudopotentials acting on the full correlated wave function.
The action of the nonlocal pseudopotential on the
wave function can be written as a sum of contributions
from each electron and each angular momentum channel. The contribution to the local energy E L
⫽⌿ T⫺1 Ĥ⌿ T made by the nonlocal pseudopotential
terms is
V nl⫽⌿ T⫺1 V̂ nl⌿ T
⫽
ps
⌿ T ⫽ 兺 V nl,i ,
兺i ⌿ T⫺1 V̂ nl,i
i
(8.3)
where for simplicity we consider the case of a single
atom placed at the origin. Using Eq. (8.2) we can write
the nonlocal contribution of electron i to the local energy as
V nl,i ⫽
兺l
⫻
l
ps
V nl,l
共ri兲
兺
m⫽⫺l
Y lm 共 ⍀ ri 兲
冕
* 共 ⍀ r⬘ 兲
Y lm
⌿ T 共 r1 ,...,ri⫺1 ,r⬘i ,ri⫹1 ,...,rN 兲
⌿ T 共 r1 ,...,ri⫺1 ,ri ,ri⫹1 ,...,rN 兲
E. Pseudopotentials in diffusion Monte Carlo
i
d⍀ r⬘ ,
i
(8.4)
where the angular integration is over the sphere passing
through the ith electron and centered on the origin.
Equation (8.4) can be simplified by choosing the z axis
along ri , noting that Y lm (0,0)⫽0 for m⫽0, and using
the definition of the spherical harmonics to give
V nl,i ⫽
ps
共ri兲
兺l V nl,l
⫻
2l⫹1
4␲
冕
P l 关 cos共 ␪ i⬘ 兲兴
⌿ T 共 r1 ,...,ri⫺1 ,ri⬘ ,ri⫹1 ,...,rN 兲
⌿ T 共 r1 ,...,ri⫺1 ,ri ,ri⫹1 ,...,rN 兲
The use of nonlocal pseudopotentials in DMC is more
problematic. If the Hamiltonian contains the nonlocal
ps
operator V̂ nl⫽ 兺 i V̂ nl,i
the approximate propagator includes matrix elements of the form 具 R兩 exp(⫺␶V̂nl) 兩 R⬘ 典 ,
which are not guaranteed to be non-negative for arbitrary R⬘ ,R, and ␶. Consequently, as the population of
walkers evolves according to the imaginary-time Schrödinger equation,
⳵ t f⫽
d⍀ r⬘ ,
i
(8.5)
where P l denotes a Legendre polynomial.
The integral over the surface of the sphere in Eq. (8.5)
must be evaluated numerically. The r⬘ dependence of
the many-body wave function is expected to have predominantly the angular momentum character of the orbitals in the determinantal part of the wave function. A
suitable integration scheme is therefore to use a quadrature rule that integrates products of spherical harmonics
exactly up to some maximum value l max . Appropriate
values of l max may be deduced on chemical grounds, but
one must remember that even in an atom the Jastrow
factor introduces higher angular momentum components than occur in the determinantal part of the wave
function. To avoid bias the orientation of the axes is
chosen randomly each time such an integral is evaluated.
Mitas, Shirley, and Ceperley (1991) have tested a number of quadrature grids for the Si atom and the Cu⫹ ion.
Taking Si as an example, a grid containing 6 points with
l max⫽3 is sufficient for reasonable accuracy, but considerably higher accuracy can be obtained using a grid containing 12 points with l max⫽5. In earlier work (Fahy,
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Wang, and Louie, 1988, 1990a) it was suggested that
each nonlocal integral should be evaluated by sampling
the quadrature grid several times using randomly oriented axes, but it has since been established that it is
more efficient to sample a higher-order rule once (Mitas,
Shirley, and Ceperley, 1991). Within a VMC calculation
it is often possible to use a low-order quadrature rule
because the error cancels over the run, but higher accuracy is required for wave-function optimization and
DMC calculations, which are biased by errors in the
nonlocal integration.
In principle the nonlocal energy should be summed
over all the ionic cores and all electrons in the system.
However, since the nonlocal potential of each ion is
short ranged, one need only sum over the few atoms
nearest to each electron. The additional angular integrations are computationally inexpensive within VMC because the local energy is not required during the oneelectron moves and need not be evaluated frequently.
Indeed, if the local energy is evaluated too frequently
the values obtained will be correlated and little or no
benefit will accrue.
1 2
共 Ĥ⫺E T 兲 ⌿ T
f
ⵜ f⫺ⵜ• 共 vD f 兲 ⫺
2
⌿T
⫹
再
冎
V̂ nl⌿ T V̂ nl⌽
⫺
f,
⌿T
⌽
(8.6)
the sign of a walker can change as the time evolves.
After a few time steps the sign becomes random and we
encounter a sign problem analogous to the fermion sign
problem discussed in Sec. III.D.2 and just as serious.
To circumvent this difficulty the so-called pseudopotential localization approximation has been introduced:
the term in curly brackets is neglected, thus making Eq.
(8.6) formally equivalent to an imaginary-time Schrödinger equation with local potentials. If ⌿ T ⬇⌿ 0 the error introduced by this approximation is small and is proportional to (⌿ T ⫺⌿ 0 ) 2 (Mitas, Shirley, and Ceperley,
1991). Because the localization approximation depends
on the accuracy of ⌿ T , it is necessary to use accurate
trial functions to ensure that the calculation is in the
low-variance/quadratic-convergence regime. Fortunately, as was explained in Sec. IV, accurate trial functions are available for simple solids. It is not easy to
establish the exact size of the error introduced by the
localization approximation since no better calculational
technique is available for solids, but comparisons with
70
Foulkes et al.: Quantum Monte Carlo simulations of solids
TABLE IV. Variational Monte Carlo of the Fe atom with decreasing number of valence electrons.
All-electron
E HF(a.u.)
E VMC(a.u.)
␴ E2
␬ / ␬ all
2
Efficiency⫽1/( ␬ ␴ E
)
Valence errors (eV)
⫺1262.444
⫺1263.20(2)
⬇50
1
0.02
0.
experiments and other calculations for small systems
show that in most cases it is smaller than the fixed-node
error.
The crucial speed-up resulting from the use of
pseudopotentials is clearly demonstrated by the example
of the iron atom (Mitas, 1994). Table IV give values of
2
the total energy; the variance of the local energy, ␴ E
2
⫽ 具 ⌿ T 兩 关 E L (R)⫺E V 兴 兩 ⌿ T 典 / 具 ⌿ T 兩 ⌿ T 典 ; the energy autocorrelation time ␬ divided by the all-electron value ␬ all
(values of the local energy calculated at closely separated times are statistically correlated; ␬ is a measure of
how long the interval between two energy measurements must be to ensure that these correlations are negligible); and, finally, the efficiency, which is proportional
2
to 1/( ␬ ␴ E
). The results were obtained using VMC but a
comparison of DMC results would be qualitatively similar. It is evident that the efficiency improves dramatically as the size of the core is increased, but that the
systematic errors introduced by the pseudopotentials
also increase. In the case of the Fe atom, if we seek an
accuracy of ⬃0.1 eV, the best compromise is the Ne
core. This comparison provides a quantitative example
but should not be taken as definitive: to some extent one
can change the quantities shown in Table IV through
improvement of the trial function, more efficient sampling, and more efficient coding.
Very recently, a new algorithm was suggested that
produces an upper bound on the energy even when the
Hamiltonian contains nonlocal operators. This new approach combines the localization approximation with
nonlocal operator sampling (ten Haaf et al., 1995) and
has proved very useful in lattice models. Unfortunately
there is a price to be paid: the algorithm has no zerovariance property (Ceperley and Mitas, 1996). This
might be an issue for calculations of realistic systems
using very accurate wave functions, when one is usually
working in the low-variance regime.
F. Alternatives to nonlocal pseudopotentials
The use of nonlocal pseudopotentials in DMC is both
problematic and computationally demanding. This has
motivated the development of alternative approaches,
the most important of which are the pseudoHamiltonian method of Bachelet, Ceperley, and Chiocchetti (1989) and the damped-core method of Hammond, Reynolds, and Lester (1988).
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Ne core
⫺123.114
⫺123.708(2)
1.54
⬇0.3
2.1
⬇0.1
Ar core
⫺21.387
⫺21.660(1)
0.16
⬇0.05
125.
⬇0.5
In the pseudo-Hamiltonian approach the core is removed and the local potential modified in the normal
way. However, instead of introducing nonlocal potentials, the kinetic-energy operator (Laplacian) is replaced
by a general second-order differential operator with a
position-dependent electron effective mass tensor. The
fundamental advantage of this construction is that the
Hamiltonian remains local. The pseudo-Hamiltonian approach is easily incorporated in both VMC and DMC,
but unfortunately it has inherent limitations (Foulkes
and Schlüter, 1990) connected with the required positive
definiteness of the effective mass tensor. The consequence is that pseudo-Hamiltonians can be constructed
only for a restricted set of elements in which nonlocal
effects are not very strong. Nevertheless, successful
simulations employing pseudo-Hamiltonians have been
carried out (Li, Ceperley, and Martin, 1991).
In the damped-core technique (Hammond, Reynolds,
and Lester, 1988), the space around an atom is divided
into core (inner) and valence (outer) regions. All the
core and valence electrons are included in the simulation, but electrons in the core regions are moved according to the VMC algorithm while electrons in the valence
regions are moved using DMC. When an electron moves
from one region to the other it experiences a smooth
transition between these two regimes. Although very
good results were obtained for atoms (Hammond, Reynolds, and Lester, 1988), this technique is difficult to apply to large systems. The reason is that the key problem,
namely the impact of the large energy fluctuations from
the core states, is diminished but not eliminated.
There is clearly a need for further development of
many-body pseudopotentials for use in QMC calculations. A step in this direction was taken by Acioli and
Ceperley (1994), who showed that the quality of a manybody pseudopotential depends on an accurate description of the one-electron, two-electron, etc., density matrices outside the core region. The most important
quantity is the one-electron density matrix, the eigenfunctions of which are the natural orbitals. One therefore has to ensure that the natural orbitals of the
pseudo-atom are correct outside the core region. This
work represents an advance in understanding, but has
not so far been developed into a practical scheme for
generating accurate pseudopotentials for heavy atoms.
71
Foulkes et al.: Quantum Monte Carlo simulations of solids
IX. PERIODIC BOUNDARY CONDITIONS
AND FINITE-SIZE ERRORS
A. Introduction
Current algorithms and computational resources permit QMC calculations on systems with up to about 1000
electrons. Although finite clusters of this size are very
unlike bulk material, periodic systems containing a few
hundred electrons are large enough to model many phenomena in condensed matter with good accuracy. The
limited system size inevitably introduces finite-size errors, which decrease as the number of electrons in the
periodically repeated region increases, so QMC simulation cells for perfect crystals usually consist of several
primitive unit cells. One can also model nonperiodic systems within this framework by using the supercell approach, which has proved very successful within
independent-particle methods.
B. Bloch’s theorem for many-body systems
Bloch’s theorem is one of the cornerstones of the
independent-particle electronic structure theory of solids (see, for example, Ashcroft and Mermin, 1976). It
defines the crystal momentum used to label the eigenstates in a periodic solid and allows the independentparticle problem to be solved by considering only a
single primitive unit cell. The reduction of the problem
to one within the primitive cell is possible only in
independent-particle theories; in many-electron QMC
simulations Bloch’s theorem does not obviate the need
to solve the Schrödinger equation over the entire simulation cell. However, Bloch wave functions have the
property that the local energy ⌿ T⫺1 Ĥ⌿ T has the full
translational symmetry of the Hamiltonian. Bloch functions are normally complex and so cannot be used in
fixed-node DMC calculations. If the Hamiltonian has
time-reversal symmetry, however, the real part of a
complex Bloch eigenfunction is also an eigenfunction
(although not in general a Bloch function) and so this is
not a serious problem. The Bloch functions themselves
can always be chosen real if the Bloch wave vector is
equal to half a reciprocal lattice vector, and the use of
such trial functions is advantageous because the corresponding fixed-node DMC energies are guaranteed to
be greater than or equal to the energy of the lowest
exact eigenstate with that wave vector (Foulkes, Hood,
and Needs, 1999). A real trial function also implies a
real local energy at every point in configuration space.
Bloch’s theorem arises from the translational symmetry of the many-body Hamiltonian, which may be written schematically as
N
N
1
Ĥ⫽⫺
ⵜ 2⫹
V 共r 兲
2 i⫽1 i i⫽1 ion i
兺
1
⫹
2
兺
N
N
1
,
兺
兺 兺⬘
R i⫽1 j⫽1 兩 ri ⫺rj ⫺Rs 兩
s
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
(9.1)
where 兵 Rs 其 is the set of translation vectors of the
simulation-cell lattice, the ionic potential V ion(r) has
(at least) the periodicity of 兵 Rs 其 , N is the number of
electrons in the simulation cell, and the prime on the j
summation indicates that the j⫽i term is omitted when
Rs ⫽0. The electron-electron term includes interactions
with the electron ‘‘images,’’ which are copies of the electrons displaced by simulation-cell lattice vectors Rs ;
these images model the potential due to the electrons
‘‘outside’’ the simulation cell. The ionic potential includes contributions from all the ions within the simulation cell and their images. Strictly, the potentials in Eq.
(9.1) are divergent and additional conditions have to be
imposed to make them well defined as discussed in Sec.
IX.C. At the moment, however, we are interested only
in the translational symmetry, for which Eq. (9.1) will
suffice.
The invariance of Ĥ under the translation of any electron by a vector in 兵 Rs 其 leads to the many-body Bloch
condition (Rajagopal, Needs, Kenny et al., 1994; Rajagopal et al., 1995):
冉
⌿ ks 共 兵 ri 其 兲 ⫽U ks 共 兵 ri 其 兲 exp iks •
N
兺 ri
i⫽1
冊
,
(9.2)
where the simulation-cell periodic part, U ks , is invariant
under the translation of any electron position by a vector in 兵 Rs 其 and is antisymmetric under particle exchange. [We have omitted the band index in Eq. (9.2)
and elsewhere in this section.] The ‘‘simulation-cell
wave vector’’ ks may always be reduced into the first
Brillouin zone of the simulation-cell reciprocal lattice. If
the simulation cell has translation vectors N 1 a1 , N 2 a2 ,
and N 3 a3 , where the N i are integers and the ai are the
primitive translation vectors of the underlying crystal
lattice, the first Brillouin zone of the primitive lattice
contains a grid of N 1 N 2 N 3 points with the same (reduced) value of ks . This grid becomes finer as the simulation cell is made larger and is analogous to the k-point
sampling grids used in independent-electron calculations. Note that the symmetries leading to the existence
of the good quantum number ks arise from the artificial
periodicity of the simulation-cell Hamiltonian; Eq. (9.2)
does not hold in a real solid.
As long as N 1 , N 2 , and N 3 are not all equal to 1, the
simulation-cell Hamiltonian has more translational symmetry than assumed so far, since it is also unchanged
when all the electrons are simultaneously translated by
any primitive lattice vector Rp . This additional symmetry is quite different from the translations of a single
electron considered in the discussion of ks and leads to a
second Bloch condition (Rajagopal, Needs, Kenny et al.,
1994; Rajagopal et al., 1995):
冉
⌿ kp 共 兵 ri 其 兲 ⫽W kp 共 兵 ri 其 兲 exp ikp •
N
冊
1
r ,
N i⫽1 i
兺
(9.3)
where W kp is invariant under the simultaneous translation of all electrons by any primitive lattice vector Rp
and is antisymmetric under particle exchange. The
‘‘primitive wave vector’’ kp , which can be reduced into
72
Foulkes et al.: Quantum Monte Carlo simulations of solids
the first Brillouin zone of the primitive reciprocal lattice,
gives the physical crystal momentum measured in experiments. The wave function can be chosen to satisfy
the simulation-cell and primitive-cell Bloch conditions
simultaneously, with ks and kp linked by the condition
that the wave functions of Eqs. (9.2) and (9.3) must be
identical if we translate all electrons by a simulation-cell
reciprocal-lattice vector (Rajagopal, Needs, Kenny et al.,
1994; Rajagopal et al., 1995). However, as long as the
simulation cell contains more than one primitive unit
cell, the two wave vectors are distinct labels of the
many-body Bloch wave function and both are required
to specify the full translational symmetry.
C. Periodic boundary conditions and Coulomb interactions
The potential energy of a macroscopic but finite system containing charges (electrons and nuclei) q i at positions li may be written as
V⫽
1
2
1
qq
i j
⫽ 兺 q iV i ,
兺i j(⫽i)
兺 兩 li ⫺l
2 i
j兩
(9.4)
where the sums extend over all charges in the system
and
V i⫽
q
兺 j
j(⫽i) 兩 li ⫺lj 兩
(9.5)
is the potential at li due to every charge except q i . The
potential energy of the QMC simulation cell is defined
by analogy with Eq. (9.4):
N
1
V 共 l1 ,l2 ,...,lN 兲 ⫽
qV ,
2 i⫽1 i i
兺
兺
兺
R
s
兺 j⫽1
兺
N
1
q i q j V 共 li ⫺lj 兲 ⫹
q 2␰ ,
2 i⫽1 i
E
兺
(j⫽i)
(9.8)
where V (r⫺lj ) is the potential at r due to a periodic
lattice of unit charges at positions lj ⫹Rs plus a cancelling uniform background, and
(9.7)
冉
␰ ⫽ lim V E共 r⫺li 兲 ⫺
r→li
where the prime on the j summation indicates that the
j⫽i term is omitted when Rs ⫽0. The potential V i depends only on the positions of the few hundred charges
in the simulation cell and is at best a poor imitation of
the potential of a real solid, which depends on the positions of approximately 1023 electrons and ions.
Note that Eq. (9.7) includes the interactions of q i with
its own images; if it did not, the copies of the simulation
cell in the infinite periodic lattice would have a net
charge and the summation would be divergent. Unfortunately, even when the self-interactions are included, the
sum in Eq. (9.7) is not absolutely convergent. Consider,
for example, the sum of contributions from simulation
cells lying entirely within a large but finite sphere. For
almost every configuration sampled in a QMC simulation, the simulation cell will have a nonzero dipole moment, and hence, from the viewpoint of macroscopic
electrostatics, the surfaces of the sphere will be covered
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
N
E
N
V i⫽
N
1
V 共 l1 ,...,lN 兲 ⫽
2 i⫽1
(9.6)
where N is the number of charges in the simulation cell
and V i is the potential felt by charge q i at position li in
an infinite periodic lattice of identical copies of the simulation cell. Thus
qj
⬘
,
兩
l
⫺
l
共
i
j ⫹Rs 兲 兩
j⫽1
with polarization charges. The resulting depolarization
fields are well known to affect the total energy per unit
cell (see, for example, Kittel, 1966, Chap. 12), even in
the limit as the sphere size goes to infinity. If, instead of
spheres, we had considered larger and larger clusters of
a different shape (perhaps cubes or ellipsoids), the depolarization fields would have been different and we
would have obtained a different potential energy. Since
the depolarization potential contribution to V i is not periodic, it is convenient to set the depolarization field to
zero. This can be thought of as surrounding the sphere
by a conducting medium and is sometimes known as
applying tin foil boundary conditions. The solution of
Poisson’s equation within the sphere with zero depolarization field is unique to within a constant, which does
not affect the Coulomb energy of Eq. (9.6) because the
simulation cell has no net charge. This recipe therefore
produces an unambiguous result, called the Ewald energy. For a full mathematical discussion of the difficulties associated with defining Coulomb potentials in periodic systems see De Leeuw, Perram, and Smith (1980); a
simpler treatment is given by Fraser et al. (1996).
The Ewald energy may be calculated in a number of
different ways (see Sec. X.D), but QMC simulations of
solids generally use the original Ewald method (Ewald,
1921; Tosi, 1964; De Leeuw, Perram, and Smith, 1980)
or one of its variants (Rajagopal and Needs, 1994; Natoli
and Ceperley, 1995). The Ewald energy is written in the
form
1
兩 r⫺li 兩
冊
(9.9)
is the potential at li due to its own images and background. The Ewald interaction V E is calculated using
the expression
V E共 r⫺li 兲 ⫽
兺
R
s
⫹
erfc关 ␬ 兩 r⫺ 共 li ⫹Rs 兲 兩 兴
␲
⫺ 2
兩 r⫺ 共 li ⫹Rs 兲 兩
␬ ⍀s
exp共 ⫺G s2 /4␬ 2 兲 iG •(r⫺l )
4␲
i ,
e s
⍀ s Gs ⫽0
G s2
兺
(9.10)
where ⍀ s is the volume of the simulation cell, 兵 Gs 其 is the
set of vectors reciprocal to 兵 Rs 其 , and ␬ is an adjustable
parameter that does not affect the value of V E and may
be chosen to optimize the convergence rates of the real
and reciprocal space sums.
Depolarization fields in real solids are almost always
negligibly small, either because the solid is unpolarized
or because the polarization charges are screened by
charged dust particles adhering to the surfaces. The
73
Foulkes et al.: Quantum Monte Carlo simulations of solids
Ewald method therefore gives the correct Hartree and
electron-ion energies in almost all cases. (These energies
are the same as those obtained using the k-space approach familiar from band-structure calculations.) The
electrons in real solids are in different places in every
unit cell, however, and so the correlations between electrons are not the same as in a periodic array of charges.
The spurious correlations built into the Ewald interaction produce large finite-size errors in the exchangecorrelation energy, the treatment of which is discussed
in the next section.
D. Coping with finite-size errors
1. Introduction
Using a finite simulation cell to model an extended
system introduces finite-size errors. These are one of the
major problems encountered in the application of accurate many-body techniques to extended systems, and it
is vital that they be treated properly.
Finite-size errors also occur in independent-particle
calculations, but their nature is subtly different from the
many-body case. Within the supercell approach, a point
defect in an otherwise perfect crystal is modeled by taking a simulation cell containing a single point defect and
repeating the cell periodically throughout space. If the
simulation cell is too small the interactions between the
defects in neighboring simulation cells produce a significant finite-size error. This type of finite-size error is
common to both independent-particle and many-body
calculations. A second source of error within
independent-particle calculations arises from inaccuracies in the integration over the Brillouin zone. This error
can also be thought of as a type of finite-size effect. Most
many-body simulations for periodic systems use only
one k point in the Brillouin zone of the simulation cell
(i.e., one value of ks ), and therefore they suffer from
significant finite-size effects of this type.
There is, however, an additional important contribution to the finite-size errors in many-body calculations of
solids. The periodic Ewald interaction used to model the
electron-electron interactions depends on the size and
shape of the simulation cell, and this produces errors
that are not encountered in density-functional calculations. Hartree-Fock theory also requires an explicit
electron-electron interaction, and using the Ewald form
produces an analogous finite-size error in the exchange
term (but not the Hartree term). In density-functional
theory, however, the exchange-correlation energy is obtained from a formula derived from DMC data that already contain an extrapolation to infinite simulation-cell
size (Ceperley and Alder, 1980). Density-functional calculations for periodic systems therefore do not contain a
finite-size error arising from the Ewald interaction.
2. Finite-size correction and extrapolation formulas
What practical schemes are available for coping with
finite-size errors in QMC or other many-body calculaRev. Mod. Phys., Vol. 73, No. 1, January 2001
tions? One idea is to use a finite-size correction formula
in which the energy for the infinite system is written as
E ⬁ ⫽E N ⫹ 共 E ⬁ ⫺E N 兲 ,
(9.11)
where the subscript denotes the system size. A QMC
calculation for the N-particle system is then performed
to obtain an accurate value for E N , and the correction
term in brackets is approximated using a much less expensive scheme that can be applied to very large systems. An alternative is to carry out QMC calculations
for a range of different system sizes, fit the results to
some chosen function of N, and attempt to extrapolate
to infinite system size. Correction and extrapolation procedures can be combined to give
E ⬁ ⯝E N ⫹ 共 E ⬁⬘ ⫺E ⬘N 兲 ⫹F 共 N 兲 ,
(9.12)
where the prime indicates that a less expensive scheme
is used and F(N) is an extrapolation function. The optimal form of F(N) depends on the method used to calculate the correction term. Extrapolation is costly because it involves more calculations and is prone to
inaccuracy because one has to perform a fit with only a
few noisy data points. In designing a correction/
extrapolation procedure one therefore tries to make the
extrapolation term as small as possible.
Ceperley and co-workers (Ceperley, 1978; Ceperley
and Alder, 1987; Tanatar and Ceperley, 1989; Kwon,
Ceperley, and Martin, 1998) have successfully used an
extrapolation technique, fitting to the formula
E DMC,⬁ ⯝E DMC,N ⫹a 共 T ⬁ ⫺T N 兲 ⫹
b
,
N
(9.13)
where a and b are parameters and T is the kinetic energy of the noninteracting electron gas. The b/N term
accounts for the finite-size effects arising from the interaction energy, and the difference of the parameter a
from unity accounts for the difference between the kinetic energies of the interacting and noninteracting systems. In general, the noninteracting gas correction can
be replaced by an independent-particle LDA or HF correction.
Fahy, Wang, and Louie (1990a) have used a finite-size
correction procedure in which the correction term is
evaluated within the LDA using an accurate Brillouinzone integration; however, this method does not include
a correction for the interaction energy. A related idea is
to average QMC results over different values of the
simulation-cell wave vector ks , which amounts to averaging over the boundary conditions on the simulation
cell. This method has been used for lattice Hamiltonians
(see, for example, Valentı́ et al., 1991; Gros, 1992) and
was used in the VMC calculations of Kralik, Delaney,
and Louie (1998) to evaluate the Compton profile of Si
(see Sec. V.F). Averaging over ks has the advantage that
results from more approximate methods, such as the
LDA, are not required, but it is expensive and does not
include a correction for the interaction energy.
In practice, the Brillouin-zone integration and interaction errors appear to be almost independent and can be
corrected for separately (Kent, Hood, et al., 1999). The
74
Foulkes et al.: Quantum Monte Carlo simulations of solids
Brillouin-zone integration errors are usually well accounted for by corrections derived from LDA results,
but these corrections do not remove the finite-size errors
in the interaction energy. Recently, Fraser et al. (1996),
Williamson et al. (1997), and Kent, Hood, et al. (1999)
have shown how the finite-size errors in the interaction
energy may be reduced by replacing the Ewald interaction by a ‘‘model periodic Coulomb’’ interaction. This
technique is very effective, reducing or even eliminating
the need for extrapolation. The model periodic Coulomb interaction is described in Sec. IX.D.4.
3. Choosing the simulation-cell wave vector
The computational cost of an LDA calculation is proportional to the number of k points sampled within the
primitive Brillouin zone. In a QMC calculation, however, the volume of the simulation cell is proportional to
the number of k points sampled within the primitive
Brillouin zone, and the computational cost increases approximately as the cube of the volume of the simulation
cell. This makes it much more difficult to reduce the
Brillouin-zone integration errors by improving the
k-point sampling. The problem may be ameliorated by
applying corrections calculated using a simpler method
such as the LDA, but one nevertheless expects to obtain
more accurate answers if the required corrections are as
small as possible. Since the Brillouin-zone integration
errors for a given simulation cell depend quite sensitively on the choice of the simulation-cell wave vector
ks , it is important to choose ks carefully.
The problem of choosing ks is entirely analogous to
that of choosing ‘‘special k points’’ (Baldereschi, 1973;
Monkhorst and Pack, 1976) in band-structure calculations, and ideas from that field can be taken over directly into many-body calculations (Rajagopal et al.,
1994; 1995; Kent, Hood, et al., 1999). The theory of special k points was first developed by Baldereschi (1973),
who defined the ‘‘mean-value point,’’ which is a k point
at which smooth periodic functions of the wave vector
accurately approximate their averages over the Brillouin
zone. However, since the Baldereschi mean-value point
is not equal to half a reciprocal lattice vector, one cannot
construct real Bloch wave functions at that point. A
more convenient and very successful approach is to
choose ks from among those points that allow real Bloch
wave functions according to the symmetrized planewave test of Brillouin-zone integration quality introduced by Baldereschi (1973). The relationship of this
approach to the widely used Monkhorst-Pack scheme of
Brillouin-zone integration has been discussed by Kent,
Hood, et al. (1999).
errors must arise from the exchange-correlation energy.
This can be written as the energy of interaction of the
electrons with their own exchange-correlation holes,
which are normally short ranged and quite insensitive to
the system size. The large finite-size effect must therefore arise from the dependence of the interaction itself
on the system size.
This conclusion can be rationalized by the following
analysis. Expanding the Ewald interaction around zero
separation gives
冉 冊
r4
1
2␲ T
V E共 r兲 ⫽ ⫹c⫹
r •D•r⫹O 5/3 ,
r
3⍀ s
⍀s
where c is a constant, ⍀ s is the volume of the simulation
cell, and the tensor D depends on the shape of the simulation cell (for a cell with cubic symmetry D⫽I). The
value of the constant c does not affect the total energy
because the simulation cell is neutral. The other deviations from 1/r make the Ewald interaction periodic, but
are also responsible for the spurious contribution to the
exchange-correlation energy (Fraser et al., 1996; Williamson et al., 1997). For cubic cells these terms are positive for small r and so the exchange-correlation energy
is more negative than it should be. Since the leading
correction is proportional to the inverse of the
simulation-cell volume, the error per electron is inversely proportional to the number of electrons in the
cell. Detailed calculations have confirmed this picture
(Fraser et al., 1996; Williamson et al., 1997; Kent, Hood,
et al., 1999).
Clearly it would be desirable to remove this spurious
contribution to the exchange-correlation energy, but we
must remember that the Hartree energy is given correctly by the Ewald interaction. The key requirements
for a model Coulomb interaction giving small
interaction-energy finite-size effects in simulations using
periodic boundary conditions are therefore: (i) it should
give the Ewald interaction for the Hartree terms, and
(ii) it should be exactly 1/r for the interaction with the
exchange-correlation hole. Unfortunately, the only periodic solution of Poisson’s equation for a periodic array
of charges is the Ewald interaction, which obeys criterion (ii) only in the limit of an infinitely large simulation
cell.
This problem was solved by Williamson et al. (1997)
by abandoning the Ewald interaction and instead using a
model periodic Coulomb interaction that satisfies both
criteria,
Ĥ e⫺e⫽
f 共 ri ⫺rj 兲
兺
i⬎j
4. Interaction-energy finite-size effects
Where does the large finite-size effect in the Ewald
interaction energy come from? Since the LDA corrections to the QMC energies are effective in removing the
finite-size errors in the noninteracting kinetic energy,
the interaction of the electrons with the external potential, and the Hartree energy, the remaining finite-size
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
(9.14)
⫹
兺i
冕
WS
关 V E共 ri ⫺r兲 ⫺f 共 ri ⫺r兲兴 n 共 r兲 dr, (9.15)
where n is the electronic number density and
f 共 r兲 ⫽
1
.
rm
(9.16)
75
Foulkes et al.: Quantum Monte Carlo simulations of solids
The definition of the cutoff Coulomb function f involves
a minimum image convention whereby the interelectron
vector r is reduced into the Wigner-Seitz cell of the
simulation-cell lattice by removal of simulation-cell lattice vectors, leaving a vector rm . This ensures that Ĥ e⫺e
has the correct translational and rotational symmetries.
The second term of Eq. (9.15) is a mean-field-like oneelectron potential that accounts for the difference between the Hartree energies calculated using the Ewald
and model periodic Coulomb interactions. As in the case
of the Hartree potential in HF and density-functional
calculations, a double-counting correction has to be
added to the expression for the total electron-electron
interaction energy to prevent this term’s being counted
twice. The electron-electron energy as calculated with
the model periodic Coulomb interaction then consists of
the sum of the Hartree energy calculated with the Ewald
interaction and the exchange-correlation energy calculated with the cutoff interaction f (Williamson et al.,
1997). The mean-field-like potential depends on the
electron density, which could in principle be calculated
self-consistently; in practice, however, the LDA (or HF)
density is normally accurate enough to serve as a good
approximation.
Tests using the model periodic Coulomb interaction
have shown that it dramatically reduces the finite-size
effects in the interaction energy (Fraser et al., 1996; Williamson et al., 1997; Kent, Hood, et al., 1999). Wave
functions generated by variance minimization using the
Ewald and model periodic Coulomb interactions are almost identical, and so properties other than the energy,
such as pair-correlation functions, are hardly affected by
the choice of interaction. As the model periodic Coulomb interaction gives the correct interaction between
the electrons at short distances, it may give a better account of, for example, the short-distance behavior of the
pair-correlation function, but this point has not yet been
tested by calculations.
desirable because of the loss of accuracy. Because we
require the single-particle orbitals at points in real space,
the most natural procedure is to use a basis that is localized in real space, so that the cost of evaluating the
single-particle orbitals will not increase rapidly with system size. Tabulating the orbitals on a grid is feasible for
small systems, but for bigger systems the storage requirements are too large. Plane-wave expansions are inefficient in most cases, but if there is a short repeat
length in the problem they can still be the best solution.
The best all round compromise appears to be to use an
expansion in Gaussian orbitals, which are reasonably
well localized in space and may be obtained from standard electronic structure packages.
B. Evaluation of the trial wave function
One of the main tasks in a QMC program is to evaluate the ratio ⌿(Rnew)/⌿(Rold) needed to calculate the
acceptance probability of a trial move from Rold to Rnew.
In our QMC calculations electrons are moved one at a
time and therefore Rold and Rnew differ in the position of
one electron only.
Consider a Slater-Jastrow wave function of the form
⌿ 共 R兲 ⫽e J(R) D ↑ 共 r1 ,...,rN ↑ 兲 D ↓ 共 rN ↑ ⫹1 ,...,rN 兲 ,
where D ↑ and D ↓ are Slater determinants for the
spin-up and spin-down electrons, respectively, and
N
J 共 R兲 ⫽
兺
i⫽1
It is very important to use accurate single-particle orbitals to form the Slater determinants in the trial wave
function. To this end we use sophisticated codes to perform the required density-functional and HF calculations such as the molecular Gaussian orbital packages
GAUSSIAN (Frisch et al., 1995) and GAMESS (Schmidt
et al., 1993), the molecular and solid CRYSTAL package
(Dovesi et al., 1996), and various plane-wave pseudopotential packages. Within a QMC calculation itself the
evaluation of the trial wave function and its derivatives
can take up to half the computing time. It is therefore
necessary to develop efficient methods to evaluate
single-particle orbitals, Slater determinants, and Jastrow
factors. A central issue is how to represent the singleparticle orbitals. This question is logically separate from
the basis used in the HF and density functional calculations, although reexpansion in a different basis set is unRev. Mod. Phys., Vol. 73, No. 1, January 2001
N
␹ ␴ i 共 ri 兲 ⫺
1
2 i⫽1
N
兺 兺
u ␴ i , ␴ j 共 ri ,rj 兲 .
(10.2)
j⫽1
(j⫽i)
How does this wave function change when the ith
to rnew
? Since the
spin-up electron is moved from rold
i
i
spin-down Slater determinant does not alter, the wavefunction ratio may be written as
X. COMPUTATIONAL ISSUES
A. Representation of the single-particle orbitals
(10.1)
⌿ 共 Rnew兲
⫽q ↑
⌿ 共 Rold兲
冋
冋
exp ␹ ↑ 共 rnew
兲⫺
i
exp
␹ ↑ 共 rold
i 兲⫺
册
册
j(⫽i)
兺
u ↑, ␴ j 共 rnew
,rj 兲
i
兺
u ↑, ␴ j 共 rold
i ,rj 兲
j(⫽i)
, (10.3)
where
q ↑⫽
D ↑ 共 r1 ,r2 ,...,rnew
,...,rN 兲
i
D ↑ 共 r1 ,r2 ,...,rold
i ,...,rN 兲
.
(10.4)
Assuming that the matrix of cofactors corresponding to
the determinant D ↑ is available, there are efficient algorithms for evaluating q ↑ in a time proportional to N. If
the trial move is accepted, the matrix of cofactors may
then be updated in a time proportional to N 2 . These
algorithms, which are well explained by Fahy, Wang,
and Louie (1990a), make use of the Sherman-Morrison
formula (see, for example, Press et al., 1992) and were
first used in QMC simulations by Ceperley, Chester, and
Kalos (1977). The contributions from the Jastrow factor
may also be evaluated in a time proportional to N.
76
Foulkes et al.: Quantum Monte Carlo simulations of solids
C. Evaluation of the local energy
D. Efficient treatment of the Coulomb interaction
Another important part of any Monte Carlo program
is the evaluation of the local energy
As explained in Sec. IX.C, the Coulomb energy of the
simulation cell is defined in terms of the Ewald potential
V E(r), which is obtained by solving Poisson’s equation
subject to periodic boundary conditions. The Ewald potential may be calculated using the original Ewald
method (Ewald, 1921; Tosi, 1964; De Leeuw, Perram,
and Smith, 1980), which was invented long before the
age of computers but is still in widespread use today.
The idea is to write V E(r) as the sum of two separate
series, one in real space and one in k space, both of
which can be made to converge fairly rapidly. In QMC
simulations the electrons are moved one at a time and so
it is necessary to calculate V E(r) at a point in space, not
just the total potential energy of the cell. This makes
other well-known methods such as the particle-particle
and particle-mesh schemes (Hockney and Eastwood,
1981) and Greengard’s multipole algorithm (Greengard
and Rokhlin, 1987) less effective. In addition, Greengard’s method is restricted to Coulomb potentials, and,
although O(N), is inefficient unless the simulation cell
contains at least several thousand particles. Consequently, the original Ewald method or variants thereof
are normally used. We favor the efficient variants developed by Rajagopal and Needs (1994) and by Natoli and
Ceperley (1995) based on an optimized division between
the real- and k-space summations.
E L 共 R兲 ⫽
Ĥ⌿ 共 R兲 T̂⌿ 共 R兲 V̂ nl⌿ 共 R兲
⫽
⫹
⫹V loc共 R兲 ,
⌿ 共 R兲
⌿ 共 R兲
⌿ 共 R兲
(10.5)
where T̂ is the kinetic-energy operator, V loc(R) is the
local part of the potential energy, and V̂ nl is the nonlocal
part arising from the pseudopotential. In DMC simulations, in particular, the local energy determines the
branching ratio and has to be evaluated after every oneelectron trial move. This is a very time-consuming process and so efficient algorithms are needed.
The evaluation of V loc(R) is in principle straightforward (although some subtleties connected with the use
of periodic boundary conditions were discussed in Sec.
IX.C). Efficient algorithms for evaluating Coulomb potentials in periodic systems are discussed in Sec. X.D.
The treatment of the nonlocal pseudopotential term was
discussed in Secs. VIII.D and VIII.E.
The kinetic contribution to the local energy may be
decomposed into contributions from each electron,
N
K⫽
N
兺 K i ⫽ i⫽1
兺 ⌿ ⫺1
i⫽1
冉
冊
1
⫺ ⵜ 2i ⌿.
2
(10.6)
Because of the exponential form of the Jastrow factor, it
is convenient to reexpress K i in terms of the logarithm
of ⌿:
K i ⫽2T i ⫺ 兩 Fi 兩 ,
2
(10.7)
where
Fi ⫽
1
&
ⵜ i 共 ln 兩 ⌿ 兩 兲 ⫽
1 ⵜ i⌿
,
& ⌿
(10.8)
冉 冊
1
1 ⵜ 2i ⌿ 1 ⵜ i ⌿
⫹
T i ⫽⫺ ⵜ 2i 共 ln 兩 ⌿ 兩 兲 ⫽⫺
4
4 ⌿
4 ⌿
2
.
(10.9)
The three-dimensional vector Fi is proportional to that
part of the 3N-dimensional drift velocity vector vD associated with electron i. The drift velocity plays an important role in the DMC algorithm, as explained in Sec.
III D. Integration by parts shows that
具 K i 典 ⫽ 具 兩 Fi 兩 2 典 ⫽ 具 T i 典 ,
(10.10)
where the angular brackets denote averages over the
probability density 兩 ⌿(R) 兩 2 . The kinetic energy may be
evaluated using any of these three estimators, but 具 K 典 is
the most useful because it has the lowest variance.
Moreover, when 具 K 典 is added to the average of the potential energy, the variance of the sum tends to zero as
the wave function approaches an eigenstate. Equation
(10.10) holds for any trial function ⌿(R) and provides a
very useful consistency check of QMC programs. T i and
Fi (and hence K i and vD ) may be evaluated efficiently
using the cofactor methods mentioned in Sec. X.B.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
E. Scaling with the number of particles
The scaling laws discussed in Sec. X.B were for a
move of a single electron. We now address the scaling of
the total computational cost with N. Consider, for example, a Monte Carlo calculation of the total energy.
2
The variance ␴ run
of the calculated average energy is
2
equal to ␴ /M, where M is the number of statistically
independent samples of E L and ␴ 2 is the variance of the
sampled values. The total computer time T run required
by the simulation is MT sample , where T sample is the time
to obtain a single statistically independent sample of
E L . Hence
T run⫽
␴ 2 T sample
.
2
␴ run
(10.11)
The local energy may always be written as a sum of
contributions from each electron, and in large enough
systems these contributions are approximately independent. The value of ␴ 2 is therefore proportional to the
number of electrons N. For T sample we can write
T sample⬀N 2 ⫹␧N 3 ,
2
(10.12)
where the N term comes from (i) evaluating all N
single-particle orbitals and their derivatives at the N
new electron positions (using localized basis functions
such as Gaussians), (ii) evaluating the Jastrow factor,
and (iii) evaluating the interaction energy. The N 3 contribution arises from the N updates of the matrix of cofactors, each of which takes a time proportional to N 2 .
Because ␧ is of the order of 10⫺4 the quadratic term
Foulkes et al.: Quantum Monte Carlo simulations of solids
dominates in practice. Remembering the factor of N
from ␴ 2 , the total computer time required to reach a
2
is therefore
given target variance ␴ run
T run⬀N 3 .
(10.13)
F. QMC on parallel computers
Monte Carlo calculations are inherently parallel in nature. At their most basic they involve the calculation of a
large set of independent random numbers and then the
averaging of a set of results produced by each of these
random numbers. Coupled with the fact that QMC calculations, especially for solids, are relatively expensive
to perform on today’s workstations, they are ideal for
parallel architecture machines, which offer two to three
orders of magnitude more computational power. In fact,
the major proportion of the work highlighted in Secs. V
and VI was carried out on massively parallel computers,
which are crucial for studying these computationally demanding problems.
It is important that any practitioner of electronic
structure theory acquire the ability to use parallel computers effectively. Simply speaking, a broad definition of
a parallel computer (encompassing machines with
hundreds/thousands of CPU’s, networked workstations,
multiprocessor workstations, etc.) is that it is a machine
with a set of processors able to work cooperatively to
solve a computational problem. Our QMC codes have
been ported to both symmetric multiprocessing and
massively parallel processing platforms. Interprocessor
communications involving sending packets of data between processors are carried out using the messagepassing interface (MPI) standard (Snir et al., 1998). Key
issues for optimal performance include tuning code on a
single node as well as optimizing the distribution of
work between nodes. Important factors such as the
proper arrangement of data to maximize memory access
to registers from primary and secondary cache and main
memory strongly influence the performance of programs, especially in large systems.
Central to any Monte Carlo computation is the efficient generation of random numbers. Although it is difficult to generate truly random numbers using a computer, many deterministic techniques exist for
generating sequences of so-called pseudorandom numbers (Press et al., 1992; Knuth, 1997) that pass most of
the obvious statistical tests for randomness. On parallel
computers, the process of generating random numbers
becomes more complicated because there are many concurrently executing tasks. Recently, Ceperley, Srinivasan, and Mascagni have developed a scalable library
for pseudorandom number generation, called SPRNG
(for scalable parallel random-number generators), which
is
freely
available
on
the
Web
(URL:
www.ncsa.uiuc.edu/Apps/SPRNG). The library consists
of a collection of generators gathered together into a
single easy-to-use package that runs on almost any computing architecture. In general, however, it is not possible to specify a set of tests sufficient to guarantee that
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
77
the random numbers in a sequence are statistically independent samples of a prescribed probability distribution.
As a result, there is always the possibility that results
obtained using a Monte Carlo simulation may be artifacts of the random-number generator used. In practice,
the best way to ensure that this is not the case is to rerun
the calculations with different generators.
The central idea in doing QMC on massively parallel
processors involves one processor orchestrating the
whole simulation, i.e., the ‘‘master-slave’’ paradigm. In
VMC, each processor independently runs a simulation
and accumulates its own set of observables; the observables from different processors are gathered, averaged,
and written out by the master processor at the end of the
run. The DMC algorithm described in Sec. III.D is also
intrinsically parallel. Since DMC simulations require an
order of magnitude more computer time than comparable VMC simulations, the argument for using parallel
computers is even more compelling. The ensemble of
walkers is distributed across the nodes of the parallel
machine and each processor carries out the various
stages of the DMC algorithm (the diffusion, drift, and
creation/annihilation of walkers) on its own subset. After all the walkers have been advanced for a block of
time steps, the accumulated mean walker energies from
all the nodes are gathered and used to update the trial
energy. It is important to load balance the algorithm by
keeping the numbers of walkers on different nodes approximately equal. Like VMC, the parallel DMC algorithm achieves almost linear scaling on machines with a
few hundred nodes.
XI. CONCLUSIONS
As the previous sections have demonstrated, QMC
methods have been through a remarkably productive
period during the last 10 or 12 years. A number of theoretical and algorithmic improvements, such as the use
of pseudopotentials in both VMC and DMC, the investigation and accurate correction of finite-size effects, and
the adoption of better optimization methods and sampling techniques, have raised the usefulness of QMC to
a new level. These advances have made possible groundbreaking applications such as the evaluation of band
gaps in insulating solids, studies of exchange-correlation
holes, simulations of the jellium surface, calculations of
the properties of defects in Si, and investigations of the
geometries of Si and C clusters.
We hope that we have been able to show that QMC is
becoming a practical and powerful alternative to more
traditional electronic structure methods. The most important characteristics and advantages of the QMC
methodology can be summarized as follows:
(a)
it gives a direct and accurate wave-function-based
treatment of quantum many-body effects;
(b) it is a very general approach, applicable to solids
and molecules and able to calculate almost any
ground-state expectation value, including energies
and static correlation functions;
78
Foulkes et al.: Quantum Monte Carlo simulations of solids
TABLE V. Comparison of methods.
Method
HF
LDA
VMC
DMC
CCSD(T)c
E corr
0
N/A
⬇85%
⬇95%
⬇75% d
E coh/bind
% errors
⬇50%
15–25%
2–10%
1–4%
10–15%d
Scaling with #
of electrons
N3
N3 a
N 3 ⫹␧N 4 b
N 3 ⫹␧N 4 b
N7
Total time
for C10
14
1
16
300
1500d
a
LDA algorithms with scaling proportional to N have been proposed.
␧⬇10⫺4 .
c
Coupled cluster with single and double substitutions, including triples noniteratively.
d
With a 6-311G* basis set; in the limit of infinite basis set and order of excitations the coupledcluster method is formally exact.
b
the N 3 scaling of the computational cost is very
favorable when compared with other correlated
wave-function methods;
(d) it has the significant computational advantages of
easily achieved scalability on parallel architectures
and low storage requirements;
(e) the DMC method does not suffer from the basis set
errors inherent in other correlated wave-function
methods;
(f) it has a track record of unique calculations: the homogeneous electron gas, superfluidity in He,
exchange-correlation holes and energies, etc.
(c)
To put the QMC methods into a broader perspective
we provide a comparison of methods in Table V. This
table compares the variational and diffusion Monte
Carlo (VMC and DMC), Hartree-Fock (HF), localdensity approximation (LDA), and coupled-cluster
(CC) approaches. The comparison, which is by no
means exhaustive, is based on results gathered from the
literature as well as on our own tests, which are mostly
for sp systems with a limited number of elements such
as C, N, H, and Si. The timings given are typical, but the
relative speeds of the different methods could easily
vary by a factor of 3 or more depending on the programs, computers, and implementations used. The following quantities are listed:
sis sets and the SD(T) level of excitations, which was the
largest affordable run for C10 . The DMC method is considerably more computationally demanding than the HF
and LDA methods, but can achieve much higher accuracy. For C10 clusters, DMC is both more accurate and
less computationally demanding than CCSD(T). Quantum Monte Carlo calculations also scale much better
with system size than CCSD(T), and therefore we believe that QMC is the best available correlated wavefunction approach for calculating the energies of solids
and reasonably large molecules.
Quantum Monte Carlo methods also have limitations
and problems, the most important of which are as follows:
(a)
it is demanding to calculate first and second derivatives of the total energy with respect to the atomic
positions, and hence to estimate interatomic forces
and force constants;
(b) the results are obtained with a statistical error bar
that decays only as the inverse square root of the
computer time;
(c) for solids the results also suffer from systematic
finite-size errors;
(d) only limited information about excited states is
available;
(e) in some cases the impact of the fixed-node approximation can bias the results.
The coupled-cluster results are quoted for 6-311G* ba-
All of these problem areas are topics of current research. For example, the calculation of derivatives using
QMC techniques is being investigated by several research groups and significant advances can be expected.
The statistical nature is of course inherent in QMC
methods, and while the development of better trial wave
functions and computers will reduce the attainable error
bars, QMC methods will remain expensive compared
with more approximate methods such as HF and LDA.
What are the main tasks in a DMC calculation? In
Table VI we outline the various tasks together with their
human and computer time demands. The table shows
that the activities most demanding of human time are
the generation of the single-particle orbitals and the optimization of the many-body wave function. This high-
the percentage of the valence correlation energy
(E corr⫽E exact⫺E HF) recovered, estimated by comparison with high-accuracy calculations for small
systems and by combining HF, correlated atomic,
and experimental data for larger systems;
(b) the errors in binding or cohesive energies;
(c) the scaling of the computational cost to calculate
the total energy with the number of electrons;
(d) the actual computational time for calculations on a
C10 cluster: the GAMESS code (Schmidt et al., 1993)
was used for the HF calculations and GAUSSIAN
(Frisch et al., 1995) for CCSD(T); the target QMC
error bar was 0.01 eV/atom and the absolute unit
of cpu time was 90 seconds on a Cray C90 processor.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
(a)
79
Foulkes et al.: Quantum Monte Carlo simulations of solids
TABLE VI. Human and computational costs of a typical DMC calculation.
Human time
10%
40%
40%
10%
Task
Choice of basis set,
pseudopotential, etc.
Calculation of HF and/or
post-HF orbitals
Many-electron trial wave
function optimization
DMC
lights the importance of integrating quantum-chemistry
and condensed-matter electronic-structure software
packages with QMC methods, and of improving optimization methods and developing more efficient functional
forms for variational wave functions. The computer time
is dominated by the DMC calculations themselves,
which, however, require little human intervention.
As Table V illustrates, VMC calculations using SlaterJastrow wave functions with ⬃30 variational parameters
can recover between 75 and 85 % of the valence correlation energy, and DMC calculations can recover
roughly 95%. The remaining 5% is the fixed-node error.
Although no exact solution to the sign problem is in
sight, it has been shown that it is possible to decrease
fixed-node errors significantly, and there is reason to
hope that in the future we may be able to do calculations
in which the fixed-node errors, although still present, are
negligible (say, less than 1% of the correlation energy).
In any case, for systems with more than a few atoms,
fixed-node DMC is already much more accurate than
any competing method.
It is now becoming apparent that the accuracy of the
VMC and DMC methods is systematic: the same accuracy pattern has been demonstrated for atoms, molecules, solids, and surfaces. This observation is perhaps
the most important outcome of recent work on QMC; a
decade ago there was no general expectation that QMC
would be systematically accurate and no clear computational evidence to justify such a conclusion.
If high accuracy is required, QMC is already the
method of choice for tackling large quantum many-body
problems. In very large systems such as solids, it is also
the only practical method based on many-body correlated wave functions, the variational principle, and the
original many-electron Schrödinger equation. It is obvious, however, that much remains to be done to make
QMC as flexible and easy to use as traditional electronic
structure methods. This will require a sizable investment
in the further development of methods and algorithms,
but we are confident that the results will justify the effort.
The largely unanticipated success of QMC simulations
of solids is beginning to change the way the electronic
structure community thinks about electronic correlation.
With its emphasis on many-electron wave functions and
probabilities, QMC has shown that it is possible to study
interacting electrons in real solids using very direct computational techniques; there is no need to resort to perRev. Mod. Phys., Vol. 73, No. 1, January 2001
Computer time
0%
10%
20%
70%
turbation theory or mean-field approximations. This realization is slowly overturning decades of accepted
wisdom.
ACKNOWLEDGMENTS
We have benefited from the exchange of ideas and
results with a large number of students, postdocs, and
other collaborators, to all of whom we offer our thanks.
The work of W.M.C.F., R.J.N., and G.R. is supported by
the Engineering and Physical Sciences Research Council
of Great Britain. W.M.C.F. has also received support
from the European Union. G.R. gratefully acknowledges financial support from Hitachi Ltd. (Japan). L.M.
is supported by the State of Illinois, U.S. Department of
Energy, and National Science Foundation funds. The
calculations reported in this review would not have been
possible without access to powerful computing resources. We thank the staff of the National Center for
Supercomputing Applications (NCSA), the University
of Cambridge High Performance Computing Facility,
the Edinburgh Parallel Computing Center, and the UK
Research Councils’ CSAR Service for their generous
help and support.
REFERENCES
Acioli, P. H., and D. M. Ceperley, 1994, J. Chem. Phys. 100,
8169.
Acioli, P. H., and D. M. Ceperley, 1996, Phys. Rev. B 54,
17 199.
Alder, B. J., D. M. Ceperley, and E. L. Pollock, 1982, Int. J.
Quantum Chem. S16, 49.
Anderson, J. B., 1975, J. Chem. Phys. 63, 1499.
Anderson, J. B., 1976, J. Chem. Phys. 65, 4121.
Anderson, J. B., 1999, Rev. Comput. Chem. 13, 133.
Ashcroft, N. W., and N. D. Mermin, 1976, Solid State Physics
(Holt Saunders, Philadelphia), p. 133.
Assaraf, R., and M. Caffarel, 1999, Phys. Rev. Lett. 83, 4682.
Bachelet, G. B., D. M. Ceperley, and M. G. B. Chiocchetti,
1989, Phys. Rev. Lett. 62, 2088.
Baer, R., M. Head-Gordon, and D. Neuhauser, 1998, J. Chem.
Phys. 109, 6219.
Baer, R., and D. Neuhauser, 2000, J. Chem. Phys. 112, 1679.
Baldereschi, A., 1973, Phys. Rev. B 7, 5212.
Ballone, P., C. J. Umrigar, and P. Delaly, 1992, Phys. Rev. B
45, 6293.
Baroni, S., and S. Moroni, 1999, Phys. Rev. Lett. 82, 4745.
80
Foulkes et al.: Quantum Monte Carlo simulations of solids
Bar-Yam, Y., and J. D. Joannopoulos, 1984, Phys. Rev. B 30,
1844.
Becke, A. D., 1988, Phys. Rev. A 38, 3098.
Becke, A. D., 1998, J. Chem. Phys. 109, 2092.
Bernu, B., D. M. Ceperley, and W. A. Lester, Jr., 1990, J.
Chem. Phys. 93, 552.
Bethe, H. A., and E. E. Salpeter, 1957, Quantum Mechanics of
One- and Two-electron Atoms (Springer, Berlin).
Blankenbecler, R., D. J. Scalapino, and R. L. Sugar, 1981,
Phys. Rev. D 24, 2278.
Bloch, F., 1928, Z. Phys. 52, 555.
Bloch, F., 1929, Z. Phys. 57, 545.
Blöchl, P. E., E. Smargiassi, R. Car, D. B. Laks, W. Andreoni,
and S. T. Pantelides, 1993, Phys. Rev. Lett. 70, 2435.
Bohm, D., and D. Pines, 1953, Phys. Rev. 92, 609.
Bowen, C., G. Sugiyama, and B. J. Alder, 1994, Phys. Rev. B
50, 14 838.
Boys, S. F., and N. C. Handy, 1969, Proc. R. Soc. London, Ser.
A 310, 63.
Burke, K., F. G. Cruz, and K. C. Lam, 1998, J. Chem. Phys.
109, 8161.
Caffarel, M., and P. Claverie, 1988, J. Chem. Phys. 88, 1100.
Callaway, J., 1957, Phys. Rev. 106, 868.
Ceperley, D. M., 1978, Phys. Rev. B 18, 3126.
Ceperley, D. M., 1986, J. Stat. Phys. 43, 815.
Ceperley, D. M., 1991, J. Stat. Phys. 63, 1237.
Ceperley, D. M., 1995a, Rev. Mod. Phys. 67, 279.
Ceperley, D. M., 1995b, in Strongly Interacting Fermions and
High-T c Superconductivity, Proceedings of the Les Houches
Summer School, Session LVI, edited by B. Douçot and J.
Zinn-Justin (Elsevier, Amsterdam), p. 427.
Ceperley, D. M., 1999, Nature (London) 397, 386.
Ceperley, D. M., and B. J. Alder, 1980, Phys. Rev. Lett. 45,
566.
Ceperley, D. M., and B. J. Alder, 1981, Physica B 108, 875.
Ceperley, D. M., and B. J. Alder, 1987, Phys. Rev. B 36, 2092.
Ceperley, D. M., and B. Bernu, 1988, J. Chem. Phys. 89, 6316.
Ceperley, D. M., G. V. Chester, and M. H. Kalos, 1977, Phys.
Rev. B 16, 3081.
Ceperley, D. M., and M. H. Kalos, 1979, in Monte Carlo Methods in Statistical Physics, 2nd ed., edited by K. Binder
(Springer, Berlin), p. 145.
Ceperley, D., M. H. Kalos, and J. L. Lebowitz, 1981, Macromolecules 14, 1472.
Ceperley, D. M., and L. Mitas, 1996, in Advances in Chemical
Physics, edited by I. Prigogine and S. A. Rice (Wiley, New
York), Vol. XCIII, p. 1.
Chelikowsky, J. R., and M. L. Cohen, 1976, Phys. Rev. B 14,
556.
Christiansen, P. A., 1988, J. Chem. Phys. 88, 4867.
Christiansen, P. A., Y. S. Lee, and K. S. Pitzer, 1979, J. Chem.
Phys. 71, 4445.
Čı́žek, J., 1969, Adv. Chem. Phys. 14, 35.
Conroy, H., 1964, J. Chem. Phys. 41, 1331.
Coulson, C. A., and I. Fisher, 1949, Philos. Mag. 40, 386.
Day, O. W., D. W. Smith, and C. Garrod, 1974, Int. J. Quantum Chem., Symp. 8, 501.
DeLeeuw, S. W., J. W. Perram, and E. R. Smith, 1980, Proc. R.
Soc. London, Ser. A 373, 27.
DePasquale,
M.
F.,
S.
M.
Rothstein,
and
J.
Vrbik, 1988, J. Chem. Phys. 89, 3629.
Diedrich, D. L., and J. B. Anderson, 1992, Science 258, 786.
Dirac, P. A. M., 1929, Proc. R. Soc. London, Ser. A 123, 714.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Donsker, M. D., and M. Kac, 1950, J. Res. Natl. Bur. Stand. 44,
551.
Dovesi, R., et al., 1996, CRYSTAL95 User’s Manual (University
of Torino, Torino, Italy).
Dreizler, R. M., and E. K. U. Gross, 1990, Density Functional
Theory (Springer, Berlin).
Eaglesham, D., 1995, Phys. World 8 (11), 41.
Eckstein, H., and W. Schattke, 1995, Physica A 216, 151.
Eckstein, H., W. Schattke, M. Reigrotzki, and R. Redmer,
1996, Phys. Rev. B 54, 5512.
Eisenberger, P., and P. M. Platzman, 1970, Phys. Rev. A 2, 415.
Engel, E., and S. H. Vosko, 1993, Phys. Rev. B 47, 13 164.
Ewald, P. P., 1921, Ann. Phys. (Leipzig) 64, 253.
Fahey, P. M., P. B. Griffin, and J. D. Plummer, 1989, Rev.
Mod. Phys. 61, 289.
Fahy, S., 1999, in Quantum Monte Carlo Methods in Physics
and Chemistry, Vol. 525 of NATO Advanced-Study Institute,
Series C: Mathematical and Physical Sciences, edited by M. P.
Nightingale and C. J. Umrigar (Kluwer Academic, Dordrecht), p. 101.
Fahy, S., X. W. Wang, and S. G. Louie, 1988, Phys. Rev. Lett.
61, 1631.
Fahy, S., X. W. Wang, and S. G. Louie, 1990a, Phys. Rev. B 42,
3503.
Fahy, S., X. W. Wang, and S. G. Louie, 1990b, Phys. Rev. Lett.
65, 1478.
Farid, B., and R. J. Needs, 1992, Phys. Rev. B 45, 1067.
Feenberg, E., 1969, Theory of Quantum Fluids (Academic,
New York).
Feller, W., 1968, An Introduction to Probability Theory and its
Applications, 3rd ed. (Wiley, New York), Vol. 1.
Feynman, R. P., and M. Cohen, 1956, Phys. Rev. Lett. 102,
1189.
Filippi, C., and D. M. Ceperley, 1999, Phys. Rev. B 59, 7909.
Filippi, C., and S. Fahy, 2000, J. Chem. Phys. 112, 3523.
Filippi, C., X. Gonze, and C. J. Umrigar, 1996, in Recent Developments and Applications of Density Functional Theory,
edited by J. M. Seminario (Elsevier, Amsterdam), p. 295.
Filippi, C., and C. J. Umrigar, 1996, J. Chem. Phys. 105, 213.
Filippi, C., C. J. Umrigar, and M. Taut, 1994, J. Chem. Phys.
100, 1290.
Flad, H.-J., M. Caffarel, and A. Savin, 1997, in Recent Advances in Quantum Monte Carlo Methods, edited by W. A.
Lester, Jr., Recent Advances in Computational Chemistry
(World Scientific, Singapore), Vol. 2, p. 73.
Fock, V., 1930, Z. Phys. 61, 126.
Foulkes, W. M. C., R. Q. Hood, and R. J. Needs, 1999, Phys.
Rev. B 60, 4558.
Foulkes, W. M. C., and M. Schlüter, 1990, Phys. Rev. B 42,
11 505.
Frank, W., U. Gösele, H. Mehrer, and A. Seeger, 1985, in Diffusion in Crystalline Solids, edited by G. E. Murch and A. S.
Nowick (Academic, Orlando, FL), p. 64.
Fraser, L. M., W. M. C. Foulkes, G. Rajagopal, R. J. Needs, S.
D. Kenny, and A. J. Williamson, 1996, Phys. Rev. B 53, 1814.
Frisch, M. J., et al., 1995, GAUSSIAN 94 Users Manual (Gaussian
Inc., Pittsburgh, PA).
Garmer, D. R., and J. B. Anderson, 1988, J. Chem. Phys. 89,
3050.
Glauser, W. A., W. R. Brown, W. A. Lester, Jr., D. Bressanini,
B. L. Hammond, and M. L. Koszykowski, 1992, J. Chem.
Phys. 97, 9200.
Foulkes et al.: Quantum Monte Carlo simulations of solids
Gösele, U., A. Plössl, and T. Y. Tan, 1996, in Process Physics
and Modeling in Semiconductor Technology, edited by G. R.
Srinivasan, C. S. Murthy, and S. T. Dunham (Electrochemical
Society, Pennington, NJ), p. 309.
Greengard, L., and V. Rokhlin, 1987, J. Comput. Phys. 73, 325.
Grimm, R. C., and R. G. Storer, 1971, J. Comput. Phys. 7, 134.
Gros, C., 1992, Z. Phys. B: Condens. Matter 86, 359.
Grossman, J. C., and L. Mitas, 1995, Phys. Rev. Lett. 74, 1323.
Grossman, J. C., and L. Mitas, 1997, Phys. Rev. Lett. 79, 4353.
Grossman, J. C., L. Mitas, and K. Raghavachari, 1995, Phys.
Rev. Lett. 75, 3870.
Gunnarsson, O., M. Jonson, and B. I. Lundqvist, 1979, Phys.
Rev. B 20, 3136.
Hamann, D. R., M. Schlüter, and C. Chiang, 1979, Phys. Rev.
Lett. 43, 1494.
Hammond, B. L., W. A. Lester, Jr., and P. J. Reynolds, 1994,
Monte Carlo Methods in Ab Initio Quantum Chemistry
(World Scientific, Singapore).
Hammond, B. L., P. J. Reynolds, and W. A. Lester, Jr., 1987, J.
Chem. Phys. 87, 1130.
Hammond, B. L., P. J. Reynolds, and W. A. Lester, Jr., 1988,
Phys. Rev. Lett. 61, 2312.
Harju, A., B. Barbiellini, S. Siljamäki, R. M. Nieminen, and G.
Ortiz, 1997, Phys. Rev. Lett. 79, 1173.
Harrison, R. J., and N. C. Handy, 1985, Chem. Phys. Lett. 113,
257.
Hartree, D. R., 1928, Proc. Cambridge Philos. Soc. 24, 89.
Heitler, W., and F. London, 1927, Z. Phys. 35, 557.
Hemley, R. J., H. K. Mao, L. W. Finger, A. P. Jephcoat, R. M.
Hazen, and C. S. Zha, 1990, Phys. Rev. B 42, 6458.
Hockney, R. W., and J. W. Eastwood, 1981, Computer Simulation Using Particles (McGraw-Hill, Maidenhead, NJ).
Hohenberg, P., and W. Kohn, 1964, Phys. Rev. 136, B864.
Hood, R. Q., M.-Y. Chou, A. J. Williamson, G. Rajagopal, R.
J. Needs, and W. M. C. Foulkes, 1997, Phys. Rev. Lett. 78,
3350.
Hood, R. Q., M.-Y. Chou, A. J. Williamson, G. Rajagopal, and
R. J. Needs, 1998, Phys. Rev. B 57, 8972.
Huang, C.-J., C. Filippi, and C. J. Umrigar, 1998, J. Chem.
Phys. 108, 8838.
Huang, C.-J., C. J. Umrigar, and M. P. Nightingale, 1997, J.
Chem. Phys. 107, 3007.
Hückel, E., 1931a, Z. Phys. 70, 204.
Hückel, E., 1931b, Z. Phys. 72, 310.
Hückel, E., 1932, Z. Phys. 76, 628.
Hund, F., 1928, Z. Phys. 51, 759.
Hurley, M. M., and P. A. Christiansen, 1987, J. Chem. Phys. 86,
1069.
Hybertsen, M. S., and S. G. Louie, 1986, Phys. Rev. B 34, 2920.
Jastrow, R. J., 1955, Phys. Rev. 98, 1479.
Jiménez, I., L. J. Terminello, D. G. J. Sutherland, J. A. Carlisle, E. L. Shirley, and F. J. Himpsel, 1997, Phys. Rev. B 56,
7215.
Jones, M. D., G. Ortiz, and D. M. Ceperley, 1997, Phys. Rev. E
55, 6202.
Jones, R. O., and O. Gunnarsson, 1989, Rev. Mod. Phys. 61,
689.
Kalos, M. H., 1962, Phys. Rev. 128, 1791.
Kalos, M. H., 1967, J. Comput. Phys. 2, 257.
Kalos, M. H., D. Levesque, and L. Verlet, 1974, Phys. Rev. A
9, 257.
Kalos, M., and F. Pederiva, 1999, in Quantum Monte Carlo
Methods in Physics and Chemistry, Vol. 525 of NATO
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
81
Advanced-Study Institute, Series C: Mathematical and Physical Sciences, edited by M. P. Nightingale and C. J. Umrigar
(Kluwer Academic, Dordrecht), p. 263.
Kalos, M. H., and P. A. Whitlock, 1986, Monte Carlo Methods
Volume 1: Basics (Wiley, New York).
Karlin S., and H. M. Taylor, 1981, A Second Course in Stochastic Processes (Academic, New York).
Kato, T., 1957, Commun. Pure Appl. Math. 10, 151.
Kenny, S. D., G. Rajagopal, and R. J. Needs, 1995, Phys. Rev.
A 51, 1898.
Kenny, S. D., G. Rajagopal, R. J. Needs, W.-K. Leung, M. J.
Godfrey, A. J. Williamson, and W. M. C. Foulkes, 1996, Phys.
Rev. Lett. 77, 1099.
Kent, P. R. C., R. Q. Hood, M. D. Towler, R. J. Needs, and G.
Rajagopal, 1998, Phys. Rev. B 57, 15 293.
Kent, P. R. C., R. Q. Hood, A. J. Williamson, R. J. Needs, W.
M. C. Foulkes, and G. Rajagopal, 1999, Phys. Rev. B 59,
1917.
Kent, P. R. C., R. J. Needs, and G. Rajagopal, 1999, Phys. Rev.
B 59, 12 344.
Kent, P. R. C., M. D. Towler, R. J. Needs, and G. Rajagopal,
2000, physics/9909037, Phys. Rev. B (to be published).
Kittel, C., 1966, Introduction to Solid State Physics, 3rd ed.
(Wiley, New York).
Knittle, E., R. Wentzcovitch, R. Jeanloz, and M. L. Cohen,
1989, Nature (London) 337, 349.
Knuth, D. E., 1997, The Art of Computer Programming, 3rd
ed. (Addison-Wesley, Reading, MA), Vol. 2, p. 61.
Kohn, W., and L. J. Sham, 1965, Phys. Rev. A 140, A1133.
Kralik, B., P. Delaney, and S. G. Louie, 1998, Phys. Rev. Lett.
80, 4253.
Krotscheck, E., W. Kohn, and G.-X. Qian, 1985, Phys. Rev. B
32, 5693.
Kwon, Y., D. M. Ceperley, and R. M. Martin, 1993, Phys. Rev.
B 48, 12 037.
Kwon, Y., D. M. Ceperley, and R. M. Martin, 1994, Phys. Rev.
B 50, 1684.
Kwon, Y., D. M. Ceperley, and R. M. Martin, 1996, Phys. Rev.
B 53, 7376.
Kwon, Y., D. M. Ceperley, and R. M. Martin, 1998, Phys. Rev.
B 58, 6800.
Lam, L., and P. M. Platzman, 1974, Phys. Rev. B 9, 5122.
Langreth, D. C., and M. J. Mehl, 1983, Phys. Rev. B 28, 1809.
Lester, W. A., 1997, Ed., Recent Advances in Quantum Monte
Carlo Methods, Recent Advances in Computational Chemistry, No. 2 (World Scientific, Singapore).
Leung, W.-K., R. J. Needs, G. Rajagopal, S. Itoh, and S. Ihara,
1999, Phys. Rev. Lett. 83, 2351.
Li, X.-P., D. M. Ceperley, and R. M. Martin, 1991, Phys. Rev.
B 44, 10 929.
Li, X.-P., R. J. Needs, R. M. Martin, and D. M. Ceperley, 1992,
Phys. Rev. B 45, 6124.
Lorenzana, H. E., I. F. Silvera, and K. A. Goettel, 1990, Phys.
Rev. Lett. 64, 1939.
Magro, W. R., D. M. Ceperley, C. Pierleoni, and B. Bernu,
1996, Phys. Rev. Lett. 76, 1240.
Malatesta, A., S. Fahy, and G. B. Bachelet, 1997, Phys. Rev. B
56, 12 201.
McMillan, W. L., 1965, Phys. Rev. 138, A442.
Meierovich, M., A. Mushinski, and M. P. Nightingale, 1996, J.
Chem. Phys. 105, 6498.
Metropolis, N., A. W. Rosenbluth, M. N. Rosenbluth, A. H.
Teller, and E. Teller, 1953, J. Chem. Phys. 21, 1087.
82
Foulkes et al.: Quantum Monte Carlo simulations of solids
Militzer, B., W. Magro, and D. M. Ceperley, 1999, Contrib.
Plasma Phys. 39, 151.
Mitas, L., 1993, in Computer Simulation Studies in Condensed
Matter Physics V, edited by D. P. Landau, K. K. Mon, and
H.-B. Schuttler (Springer, Berlin), p. 94.
Mitas, L., 1994, Phys. Rev. A 49, 4411.
Mitas, L., 1996, Comput. Phys. Commun. 96, 107.
Mitas, L., and J. C. Grossman, 1997, in Recent Advances in
Quantum Monte Carlo Methods, Recent Advances in Computational Chemistry, edited by W. A. Lester, Jr. (World Scientific, Singapore), Vol. 2, p. 133.
Mitas, L., and R. M. Martin, 1994, Phys. Rev. Lett. 72, 2438.
Mitas, L., E. L. Shirley, and D. M. Ceperley, 1991, J. Chem.
Phys. 95, 3467.
Monkhorst, H. J., and J. D. Pack, 1976, Phys. Rev. B 13, 5188.
Moroni, S., D. M. Ceperley, and G. Senatore, 1992, Phys. Rev.
Lett. 69, 1837.
Moroni, S., D. M. Ceperley, and G. Senatore, 1995, Phys. Rev.
Lett. 75, 689.
Morrell, M. M., R. G. Parr, and M. Levy, 1975, J. Chem. Phys.
62, 549.
Moskowitz, J. W., K. E. Schmidt, M. A. Lee, and M. H. Kalos,
1982, J. Chem. Phys. 77, 349.
Müller, W., J. Flesch, and W. Meyer, 1984, J. Chem. Phys. 80,
3297.
Müller, W., and W. Meyer, 1984, J. Chem. Phys. 80, 3311.
Mulliken, R. S., 1928, Phys. Rev. 32, 186.
Murphy, R. B., and R. A. Friesner, 1998, Chem. Phys. Lett.
288, 403.
Natoli, V., and D. M. Ceperley, 1995, J. Comput. Phys. 117,
171.
Natoli, V., R. M. Martin, and D. M. Ceperley, 1993, Phys. Rev.
Lett. 70, 1952.
Natoli, V., R. M. Martin, and D. M. Ceperley, 1995, Phys. Rev.
Lett. 74, 1601.
Nekovee, M., W. M. C. Foulkes, and R. J. Needs, 2000,
cond-mat/0008181.
Nekovee, M., W. M. C. Foulkes, A. J. Williamson, G. Rajagopal, and R. J. Needs, 1999, Adv. Quantum Chem. 33, 189.
Nightingale, M. P., 1999, in Quantum Monte Carlo Methods in
Physics and Chemistry, Vol. 525 of NATO Advanced-Study
Institute, Series C: Mathematical and Physical Sciences edited
by M. P. Nightingale and C. J. Umrigar (Kluwer Academic,
Dordrecht), p. 1.
Nightingale, M. P., and C. J. Umrigar, 1997, in Recent Advances in Quantum Monte Carlo Methods, Recent Advances
in Computational Chemistry, edited by W. A. Lester, Jr.
(World Scientific, Singapore), Vol. 2, p 201.
Nightingale, M. P., and C. J. Umrigar, 1999, Eds., Quantum
Monte Carlo Methods in Physics and Chemistry, Vol. 525 of
NATO Advanced-Study Institute, Series C: Mathematical and
Physical Sciences (Kluwer Academic, Dordrecht).
Ortiz, G., and P. Ballone, 1994, Phys. Rev. B 50, 1391.
Ortiz, G., D. M. Ceperley, and R. M. Martin, 1993, Phys. Rev.
Lett. 71, 2777.
Ortiz, G., M. Harris, and P. Ballone, 1999, Phys. Rev. Lett. 82,
5317.
Pack, R. T., and W. B. Brown, 1966, J. Chem. Phys. 45, 556.
Pandey, K. C., 1986, Phys. Rev. Lett. 57, 2287.
Panoff, R. M., and J. Carlson, 1989, Phys. Rev. Lett. 62, 1130.
Parr, R. G., and W. Yang, 1989, Density-Functional Theory of
Atoms and Molecules (Oxford University, New York).
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
Perdew, J. P., J. A. Chevary, S. H. Vosko, K. A. Jackson, M. R.
Pederson, D. J. Singh, and C. Fiolhais, 1992, Phys. Rev. B 46,
6671; 48, 4978(E).
Perdew, J. P., and A. Zunger, 1981, Phys. Rev. B 23, 5048.
Pierleoni, C., B. Bernu, D. M. Ceperley, and W. R. Magro,
1994, Phys. Rev. Lett. 73, 2145.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, and B. P.
Flannery, 1992, Numerical Recipes: The Art of Scientific Computing, 2nd ed. (Cambridge University, Cambridge, England).
Raghavachari, K., D. L. Strout, G. K. Odom, G. E. Scuseria, J.
A. Pople, B. G. Johnson, and P. M. W. Gill, 1993, Chem.
Phys. Lett. 214, 357.
Rajagopal, G., and R. J. Needs, 1994, J. Comput. Phys. 115,
399.
Rajagopal, G., R. J. Needs, A. James, S. D. Kenny, and W. M.
C. Foulkes, 1995, Phys. Rev. B 51, 10 591.
Rajagopal, G., R. J. Needs, S. Kenny, W. M. C. Foulkes, and
A. James, 1994, Phys. Rev. Lett. 73, 1959.
Reynolds, P. J., D. M. Ceperley, B. J. Alder, and W. A. Lester,
Jr., 1982, J. Chem. Phys. 77, 5593.
Rohlfing, M., P. Krüger, and J. Pollmann, 1993, Phys. Rev. B
48, 17 791.
Sakurai, Y., Y. Tanaka, A. Bansil, S. Kaprzyk, A. T. Stewart,
Y. Nagashima, T. Hydo, S. Nanao, H. Kawata, and N. Shiotani, 1995, Phys. Rev. Lett. 74, 2252.
Schmidt, K. E., M. A. Lee, M. H. Kalos, and G. V. Chester,
1981, Phys. Rev. Lett. 47, 807.
Schmidt, K. E., and J. W. Moskowitz, 1990, J. Chem. Phys. 93,
4172.
Schmidt, K. E., and J. W. Moskowitz, 1992, J. Chem. Phys. 97,
3382.
Schmidt, M. W., et al., 1993, J. Comput. Chem. 14, 1347.
Senatore, G., and N. H. March, 1994, Rev. Mod. Phys. 66, 445.
Senatore, G., S. Moroni, and D. M. Ceperley, 1999, in Quantum Monte Carlo Methods in Physics and Chemistry, Vol. 525
of NATO Advanced-Study Institute, Series C: Mathematical
and Physical Sciences, edited by M. P. Nightingale and C. J.
Umrigar (Kluwer Academic, Dordrecht), p. 183.
Shirley, E. L., and R. M. Martin, 1993, Phys. Rev. B 47, 15 413.
Shirley, E. L., L. Mitas, and R. M. Martin, 1991, Phys. Rev. B
44, 3395.
Singwi, K. S., and M. P. Tosi, 1981, in Solid State Physics,
edited by H. Ehrenreich, F. Seitz, and D. Turnbull (Academic, New York), Vol. 36, p. 177.
Slater, J. C., 1930, Phys. Rev. 35, 210.
Snir, M., S. Otto, S. Huss-Lederman, D. Walker, and J. Dongarra, 1998, MPI–The Complete Reference, 2nd ed. (MIT,
Cambridge, MA).
Sommerfeld, A., and H. Bethe, 1933, Handbuch der Physik
(Springer, Berlin), Vol. 24.
Stedman, M. L., W. M. C. Foulkes, and M. Nekovee, 1998, J.
Chem. Phys. 109, 2630.
Szabo, A., and N. S. Ostlund, 1989, Modern Quantum Chemistry (McGraw-Hill, New York).
Tanaka, S., 1993, J. Phys. Soc. Jpn. 62, 2112.
Tanaka, S., 1995, J. Phys. Soc. Jpn. 64, 4270.
Tanatar, B., and D. M. Ceperley, 1989, Phys. Rev. B 39, 5005.
ten Haaf, D. F. B., H. J. M. van Bemmel, J. M. J. van Leeuwen,
W. van Saarlos, and D. M. Ceperley, 1995, Phys. Rev. B 51,
13 039.
Tosi, M. P., 1964, in Solid State Physics, edited by H. Ehrenreich and D. Turnbull (Academic, New York), Vol. 16, p. 1.
Foulkes et al.: Quantum Monte Carlo simulations of solids
Towler, M. D., R. Q. Hood, and R. J. Needs, 2000, Phys. Rev.
B 62, 2330.
Umrigar, C. J., 1993, Phys. Rev. Lett. 71, 408.
Umrigar, C. J., and X. Gonze, 1994, Phys. Rev. A 50, 3827.
Umrigar, C. J., M. P. Nightingale, and K. J. Runge, 1993, J.
Chem. Phys. 99, 2865.
Umrigar, C. J., K. G. Wilson, and J. W. Wilkins, 1988, Phys.
Rev. Lett. 60, 1719.
Valentı́, R., C. Gros, P. J. Hirschfeld, and W. Stephan, 1991,
Phys. Rev. B 44, 13 203.
Vosko, S. H., L. Wilk, and M. Nusair, 1980, Can. J. Phys. 58,
1200.
Wigner, E., 1934, Phys. Rev. 46, 1002.
Williamson, A. J., R. Q. Hood, R. J. Needs, and G. Rajagopal,
1998, Phys. Rev. B 57, 12 140.
Rev. Mod. Phys., Vol. 73, No. 1, January 2001
83
Williamson, A. J., S. D. Kenny, G. Rajagopal, A. J. James, R.
J. Needs, L. M. Fraser, W. M. C. Foulkes, and P. Maccallum,
1996, Phys. Rev. B 53, 9640.
Williamson, A. J., G. Rajagopal, R. J. Needs, L. M. Fraser, W.
M. C. Foulkes, Y. Wang, and M.-Y. Chou, 1997, Phys. Rev. B
55, R4851.
Wu, Y. S. M., A. Kuppermann, and J. B. Anderson, 1999,
Phys. Chem. Chem. Phys. 1, 929.
Yan, Z., J. P. Perdew, S. Kurth, C. Fiolhais, and L. Almeida,
2000, Phys. Rev. B 61, 2595.
Young, D. P., D. Hall, M. E. Torelli, Z. Fisk, J. L. Sarrao, J. D.
Thompson, H. R. Ott, S. B. Oseroff, R. G. Goodrich, and R.
Zysler, 1999, Nature (London) 397, 412.
Zong, F. H., and D. M. Ceperley, 1998, Phys. Rev. E 58, 5123.